Interventions · countable events

When a commit leaves its lane, extra work triggers — automatically

Every commit here gets a deterministic, LLM-free drift receipt (the mechanism). When a receipt lands out of lane (off-lane share above the 15% threshold), that is one countable event — and it triggers an intervention: a sensemaking pass explains the drift, the fixes get checklisted, and the story is published here. Countable events are what make a lane priceable; this page is the count. Each entry is a structural alert, not a scolding — a diagnosis, not a verdict — and the ledger you are reading is a loss ratio, published by the house that carries it. Stated exceptions are what surplus looks like in print.

Out-of-lane receipts per day

The interventions — why · fix · checklist

1. 2026-09-05 · 📐 VS Code Architect · offPct 16 · structural

   lab(R4): the loss is placement 23.5 to 1 — and the split lives in the tool, not in a query

   Why it left the lane: The entire diff is inside scripts/lab/tape-steer.mjs, adding selection_trace-based fields (candidates, rankLoss, placeLoss, noTrace, ratio_placement_to_ranking) to a routing-measurement tool. None of vscode-architect.html's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) intersects scripts/lab/** at all — confirmed by reading vscode-architect.html's own 'Owned file/dir surface' list, which explicitly assigns adjacent surfaces (src/app to Builder, patents to Vault, chapters to Voice) but never scripts/lab/. So 100% of the reality mass this commit lit lands in laboratory's domain while the receipt scored it against architect's lane, because Originating-Terminal (physically VS Code) — not the commit's own Relevant-Rooms trailer (which correctly names laboratory first) — is what the receipt used as the intent lane.

   The semantic pull: Diff vocabulary — 'selection_trace', 'rec.candidates', 'PLACEMENT loss (never a candidate)', 'RANKING loss (scored, ranked out)', 'ratio_placement_to_ranking', 'rules_chosen' — is router/injection-instrumentation vocabulary, structurally laboratory territory, with zero overlap to architect's strategy/roadmap/CLAUDE.md vocabulary. No region narrative is cached for this sha, so I cannot cite specific [row,col] lattice coordinates; the pull is inferable only from the file-glob mismatch, not from a cached encircled-region readout.

   Intent–reality gap: The commit's Originating-Terminal trailer claims 📐 VS Code Architect as the lane, but the receipt measured 100% of the diff's mass (scripts/lab/tape-steer.mjs) sitting entirely inside laboratory's owned surface — a gap the commit's own message partially discloses by naming 🧪 laboratory first in Relevant-Rooms, which the receipt's lane-selection never consulted.

   • Implement scripts/room-drift-architect-check.sh per its own already-written spec in vscode-architect.html:635, including the Relevant-Rooms exemption above
   • Implement scripts/room-drift-cursor-check.sh per its mirrored spec in cursor-laboratory.html:736
   • Close sp-520/sp-521 (vscode-architect.html, due 2026-09-06) and sp-523 (cursor-laboratory.html, due 2026-09-06) — they already name this exact fix (out-of-lane receipt guard + globs-match check), don't duplicate a fourth self-prompt for it
   • Wire the guard as a pre-commit warn-only check reading `git diff --cached --name-only` against the room's owned globs, exempted by Relevant-Rooms as above
   • Audit the room's other 32 out-of-lane events in the last 14 days for the same scripts/lab-committed-from-VS-Code shape before assuming novel causes for each

   What we learned: The guard that would have caught this was already spec'd by name in both rooms' own HTML months before this event (scripts/room-drift-architect-check.sh, scripts/room-drift-cursor-check.sh) and never implemented — a documented-but-unbuilt guard is functionally identical to no guard, and the repo already has due self-prompts (sp-520, sp-521, sp-523, all 2026-09-06) chasing the same gap independently of this intervention.

   The missing guard (the deliverable): scripts/room-drift-architect-check.sh (and its cursor-laboratory sibling) — both already spec'd verbatim in .workflow/rooms/vscode-architect.html and cursor-laboratory.html, neither implemented. Must read the Relevant-Rooms trailer as an exemption, or it will false-positive-block every legitimate cross-room commit CLAUDE.md's own trailer rule sanctions.

   Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. litIntent (133) tracking walksIntent (133) and litReality (137) tracking walksReality (137) almost 1:1, with walkTruncated=false, indicates the walk is right-sized here — neither shallow (would show lit << walks) nor saturated/runaway. One concrete adjustment: when Relevant-Rooms names a room other than Originating-Terminal, re-seed a second walk from that named room's anchors and report both — turning the placement-vs-ranking distinction R4 itself introduced into a lane-attribution check, not just a routing-loss check. Experiments: Re-run tape-steer.mjs's own loadReceipts()/main() over the last 33 architect out-of-lane events and bucket each by whether Originating-Terminal matched the diff's actual owned-surface room vs. matched only Relevant-Rooms — measurable outcome: what fraction of the 33 is this exact mistagging shape · Recompute this commit's offPct using laboratory's owned globs as the intent lane instead of architect's — measurable outcome: does offPct drop below the 15 threshold, confirming the lane-selection hypothesis · Ablation-test the compression_dictionary_adds below against the last 10 architect-room commits touching scripts/lab/** and compare rendered offPct before/after

   Measured outcome (deterministic, not self-graded): room mean offPct 23→32 over 3 commits since · dictionary ablation saved 1046B

   This event inflates architect's apparent loss-ratio for a defect that isn't there — correctly-placed laboratory code committed from a documented cross-room terminal — and the premium normalizes the moment the drift receipt reads Relevant-Rooms (or the spec'd-but-unbuilt room-drift-*-check.sh guards ship) so the room's true out-of-lane rate reflects only genuine drift, not attribution noise.

   Open the signed attestation →

2. 2026-09-05 · 📐 VS Code Architect · offPct 24 · mistagged

   fix(loop): the receipt said the loop was dead while it worked — and the panels have two regimes

   Why it left the lane: No region narrative was cached for this sha, so the read has to run on file-glob + telemetry evidence instead of lattice coordinates. Of the 635 inserted lines, 501 (79%) are scripts/lab/receipt-corpus-audit.mjs (361 new lines) plus tests/lab/receipt-corpus-audit.test.mjs (140 new lines) — a statistics/corpus-aggregation module (parsePanel, loadCommits, pearson, spearman, volumeConfoundVerdict, RX_SIGMA, RX_OFFLANE) that is explicitly laboratory's territory per the commit's own Story trailer ('laboratory owns the corpus audit'). Architect's declared owned surface — docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md — is touched by zero files in this diff. litReality=135 against litIntent=109 (26 extra anchors, ~24% more reality mass than intent predicted) is the walk's fingerprint of that mismatch, and prewalkOverlap=6 (5.5% of litIntent) shows the reef itself found almost no shared vocabulary between the architect-declared intent lexicon and the laboratory-flavored reality lexicon before the walk started.

   The semantic pull: scripts/lab/receipt-corpus-audit.mjs + tests/lab/receipt-corpus-audit.test.mjs = 501/635 inserted lines under a 📐architect Originating-Terminal, despite the diff naming zero files inside architect's owned globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md). The only file that does sit in a plausibly-architect lane is scripts/ops/initiative-execute.mjs (+33/-2 lines, ~6% of the diff) — the piece the Story trailer actually assigns to architect ('architect owns the executor fix'). prewalkOverlap=6 anchors is the reef-level trace of that lexical gap between the claimed lane and the bulk of the work.

   Intent–reality gap: The commit message is unusually honest about its own multi-room composition ('laboratory owns the corpus audit... operator owns the consequence... architect owns the executor fix'), but ships as one commit under a single 📐architect Originating-Terminal. The receipt reads the diff's actual mass — 79% laboratory-flavored corpus-audit code — against a lane declared by the terminal that authored only the remaining 21%.

   • CLAUDE.md ITERATION DISCIPLINE: add the multi-room-Story-trailer-implies-split clause (see spec_rewrite).
   • data/rooms.json / .workflow/rooms/laboratory.html: add scripts/lab/** and tests/lab/** to 🧪 laboratory's owned-surface glob list — the corpus-audit module has no owned-surface entry anywhere today.
   • data/rooms.json / .workflow/rooms/architect.html: state explicitly that architect's owned surface is docs/strategy-only, so an infra/ops fix committed from this terminal should carry 🎩 operator or its actual room as Originating-Terminal, not architect by default.
   • scripts/lab/receipt-corpus-audit.mjs header comment: name the owning room (🧪 laboratory) alongside the existing FORECLOSES: line, so a future rules-for.mjs query against this path surfaces the ownership fact.
   • Extend receipt-corpus-audit.mjs's own aggregation (it already computes offlane_histogram and volumeConfoundVerdict) to add a third stratum: offPct broken down by whether a commit's Story trailer names >1 room — this event is itself a data point for that exact question.

   What we learned: A commit's Story trailer can correctly diagnose its own multi-room composition in prose and still ship as a single-room commit — the honesty of the message does not prevent the bundling the message describes. The room-tagging failure and the ITERATION DISCIPLINE failure are the same event seen from two rules, and neither rule currently reads the other's evidence.

   The missing guard (the deliverable): A warn-only pre-commit check (same class as scripts/ops/rules-for.mjs) that computes, per staged commit, the file-glob-to-room mapping of the diff's line mass and flags when the dominant room (by changed lines) differs from the commit's own Originating-Terminal trailer AND the Story trailer itself names a different room as the owner of that mass — a 'commit-room-mismatch' guard.

   Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. Right-sized, not shallow or saturated in the pathological sense (walkTruncated=false, walksIntent==litIntent, walksReality==litReality — every walked anchor lit, nothing cut off), but asymmetric: litReality sits at 135/144 (~94% of the full lattice), close to the ceiling this same commit's own R18 finding calls a saturation regime. Concrete adjustment: cap the reality-side walk's effective ply-depth to track the intent-side ply-depth on a per-commit basis, so a wordy, multi-topic commit message cannot let reality outrun intent by 26 anchors regardless of how much reality code exists. Experiments: Split this sha's own diff into its two file-glob halves (scripts/lab/*+tests/lab/* vs scripts/ops/*+tests/ops/*) and re-run receipt-corpus-audit-style measurement on each half independently — if per-half offPct drops under 15, that is a direct, testable confirmation that commit-splitting fixes the receipt. · Ablate the repeated header-comment boilerplate (the FORECLOSES:/AXIOM 1/W2 style preamble now standard across new scripts) from the reality corpus and re-measure offPct on the same commit — if unchanged, this boilerplate is reef noise rather than signal and is a safe compression-dictionary candidate. · Add a per-file-glob prewalkOverlap breakdown to the deterministic receipt row (today it reports one aggregate prewalkOverlap=6) so a mixed-room commit's overlap-per-glob is visible without re-running the whole audit.

   Measured outcome (deterministic, not self-graded): room mean offPct 16→26 over 8 commits since · dictionary ablation saved 1247B

   An out-of-lane receipt this large on a well-tested, well-documented commit is a labeling loss, not a quality loss — it inflates this room's apparent drift rate and erodes trust in the offPct signal itself; premium is restored by fixing the room-tagging convention (split-commit-per-Story-owner) rather than by re-scoping the actual code.

   Open the signed attestation →

3. 2026-09-05 · 📐 VS Code Architect · offPct 36 · structural

   fix(gem): the corpus was never blocked — the doctor asked one transport of two

   Why it left the lane: Every changed path in this commit sits outside 📐architect's declared surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md) — the diff is 100% scripts/chat/{browser-lease.sh,gem-doctor.mjs,gem-safari-transport.sh}, scripts/elicit/loop-owed.mjs, scripts/ops/initiative-loop.sh and their tests. Terms doing the pulling: `safari_gemini_tab()`, `transport · chrome cdp`, `transport · safari apple events`, `CORPUS_STALE_DAYS` — all corpus/pulse operational vocabulary. But checking where that vocabulary is actually claimed: I grepped `.workflow/rooms/*.html` for `scripts/chat` and `scripts/elicit` and got zero hits across all nine room files, including 🧭rio-navigator.html (whose own owned surface is `scripts/open-room-session.sh`, `.thetacog/cache/room-runners/*.sh`, the openclaw scheduler — not the chat/elicit tree the commit's Story attributes to it). The mass didn't drift toward a wrong room's lane; it drifted into surface no room has claimed at all.

   The semantic pull: The commit's own Story trailer asserts 'navigator owns the corpus and the pulse... architect owns the doctor' — a plausible-sounding split that neither .workflow/rooms/rio-navigator.html nor vscode-architect.html actually encodes. The Originating-Terminal trailer (📐 VS Code Architect) is set by physical terminal detection (open-room-session.sh --print-label), and that mechanical rule is what pulled the receipt's intent-lane onto architect's narrow docs/strategy surface regardless of what the diff actually contained.

   Intent–reality gap: The commit claims Originating-Terminal 📐 architect and Relevant-Rooms 🧭navigator/📐architect/🎩operator, implying the corpus-pipeline work has a legitimate room home; the receipt measured 36% of lit reality mass outside architect's lane because none of the three named rooms' *declared* owned-surface globs actually cover scripts/chat/ or scripts/elicit/ — the ownership claim in prose and the ownership claim in the machine-readable room files disagree.

   • Add to .workflow/rooms/rio-navigator.html's 'Owned file/dir surface': `scripts/chat/**`, `scripts/elicit/**`, `tests/chat/**`, `tests/elicit/**` — architect legitimately owns editing this file per its own charter line '.workflow/rooms/*.html — room architecture itself', so this is in-lane work for architect even though the corpus code itself is not
   • Audit the other 17 out-of-lane events in this room's last-14-days window: how many touch scripts/chat/ or scripts/elicit/ specifically — if most do, this single glob addition retroactively explains the bulk of the room's persistent offPct problem rather than 18 separate content defects
   • Check whether any other unclaimed high-traffic surface exists the same way — grep `.workflow/rooms/*.html` for `data/messages/`, `.thetacog/elicit/`, `.thetacog/pmu/` and confirm each has exactly one declared owner
   • Update the drift-receipt's lane-resolution code (wherever offPct is computed against a room) to prefer Relevant-Rooms-glob-union over Originating-Terminal-only, per the spec_rewrite above
   • File a decisions-ledger row noting the room-map gap was structural, not a per-commit tagging error, so future architect-room out-of-lane events touching scripts/chat aren't re-diagnosed from scratch

   What we learned: The room ownership map (.workflow/rooms/*.html) has at least one large, actively-developed surface — the entire message/corpus pipeline (scripts/chat/, scripts/elicit/, tests/chat/, tests/elicit/) — with NO declared owner across any of the nine rooms, despite CLAUDE.md's own THE MESSAGE SPINE entry treating it as a first-class, canonical subsystem. A commit's prose can correctly narrate room ownership ('navigator owns the corpus') while the machine-readable glob that the receipt actually measures against says nothing of the kind — prose and config can silently diverge on OWNERSHIP the same way they diverge on CLAIM elsewhere in this repo.

   The missing guard (the deliverable): A guard (e.g. tests/ops/room-surface-coverage.test.mjs) that walks every tracked path under scripts/, tests/, docs/, data/ still active in the last N days of git log and asserts it matches at least one room's 'Owned file/dir surface' glob in .workflow/rooms/*.html — flagging any actively-churning directory with zero claimants, the same class of gap this event exposed for scripts/chat and scripts/elicit.

   Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Right-sized, not shallow or saturated — fully lit with no truncation on either side. One adjustment worth testing: rerun with --max-depth+1 to confirm the 137-cell reality count is stable (not an artifact of hitting the current depth ceiling exactly), since a depth-sensitive count would change how much confidence to put in the 36% figure. Experiments: Recompute this commit's offPct against the union of Relevant-Rooms' owned-surface globs (navigator+architect+operator) instead of architect-only, and diff the resulting offPct against the shipped 36% — direct test of the structural-gap hypothesis · After adding scripts/chat/**+scripts/elicit/** to rio-navigator.html, re-render the triptych for sha 5acc8823c6 under the corrected room map and confirm offPct drops toward the 15% threshold · Seed the reef dictionary with the literal phrase 'standing in for a construct' (present in this diff and in CLAAUDE.md's PROXY IS NOT THE THING section) and measure whether prewalkOverlap rises for commits that legitimately share that governance vocabulary across docs and code comments

   Measured outcome (deterministic, not self-graded): room mean offPct 17→22 over 6 commits since · dictionary ablation saved 1486B

   An out-of-lane event caused by a genuine room-map gap (no owner for scripts/chat/scripts/elicit) reads identically to careless scope creep in the loss record unless the gap is closed in config — restoring this lane's premium means shipping the navigator glob addition and the coverage guard in the same pass, not just narrating the distinction here.

   Open the signed attestation →

4. 2026-09-05 · 🎤 voice · offPct 17

   feat(iamfim): the staged geo reveal — show them our wrong guess, then ask

   Why it left the lane: (sensemaking unavailable — facts-only intervention)

   Measured outcome (deterministic, not self-graded): room mean offPct 28→16 over 8 commits since

   Open the signed attestation →

5. 2026-09-05 · 🎤 voice · offPct 28 · structural

   feat(signups): nine "signups" in one morning, and not one of them can name the surface that produced it

   Why it left the lane: The diff's real vocabulary — `SELECT ... FROM subscribers`, `email_events`, `better-sqlite3`, `classify()`, `isGatewayBurst`, `nominator_email`, `consent audit`, `has_opted_in` — is CRM/analytics prose with zero overlap with 🎤 voice's owned lane (book chapters, blog MDX, six-needs, decentralized-talk slides). prewalkOverlap=7 (measured BEFORE the walk ran) already shows almost no shared vocabulary between this commit's reef seed and voice's intent corpus, and the commit's own 'Story' trailer explicitly places the work at 'C.Operations x A1.Strategy.Law' — an Operations coordinate, not a voice one. The walk then saturated symmetrically (litIntent=litReality=138, walkTruncated=true), meaning the mismatch baked in at seed time and the walk just propagated it.

   Intent–reality gap: The commit's own Relevant-Rooms trailer correctly names 🎩 operator, 🔨 builder, ☕ network — never 🎤 voice — and the Story trailer correctly places the work at C.Operations x A1.Strategy.Law. But the receipt still scored this commit against 🎤 voice's lane, which means the receipt pipeline is reading Originating-Terminal (which terminal happened to be open) rather than the commit's own, already-correct Relevant-Rooms attribution when it picks which lane's coordinate space to check reality against.

   • Add scripts/bookclub/** + tests/bookclub/** to 🔨 builder's (or 🎩 operator's) owned-surface glob in .workflow/rooms/*.html
   • Locate and audit the receipt pipeline's room-resolution step to confirm it uses Originating-Terminal rather than the glob-computed Relevant-Rooms trailer as fallback, and patch the preference order
   • Run node scripts/ops/rules-for.mjs scripts/bookclub/signup-provenance.mjs to confirm no CLAUDE.md placement rule for scripts/bookclub/ was missed before this becomes a repeated pattern
   • Sample this room's other 101 out-of-lane events from the last 14 days for the same 'unowned-surface + terminal-fallback' shape vs genuine voice-content drift, since the fix should only be claimed general if the pattern repeats
   • Add a firing-test guard asserting a commit whose changed files match no room's glob is NOT scored against Originating-Terminal's lane when Relevant-Rooms names a different room

   What we learned: The drift receipt's room/lane attribution can silently fall back to Originating-Terminal (whichever terminal the operator physically had open) for files that match NO room's owned-surface glob — even when the commit's own Relevant-Rooms trailer, computed by real glob-ownership matching, already names the correct room. So infra/ops scripts authored from inside a content room's terminal get scored as that content room's drift, inflating a content lane's apparent loss with work it never claimed.

   The missing guard (the deliverable): A test asserting: for a commit whose changed files match no room's owned-surface glob, the receipt's assigned room equals the highest-scoring room named in the commit's own Relevant-Rooms trailer, never the Originating-Terminal room when the two diverge.

   Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent and litReality tied exactly at 138 under walkTruncated=true, which means both sides hit the truncation ceiling rather than a natural equilibrium — that saturation conceals whether reality would diverge further from intent given more budget. Concrete adjustment: re-run this commit with a raised --max-depth/lit-ceiling; if the counts still tie, that confirms genuine saturation rather than a truncation artifact masking the true offPct. Experiments: Re-render 73920e3269's receipt against 🔨 builder's lane instead of 🎤 voice's and confirm offPct drops below the 15 threshold — a direct, cheap falsification of the mistagged-lane hypothesis · Ablation: for the next 20 commits where Originating-Terminal and Relevant-Rooms diverge, seed the reef placement from Relevant-Rooms instead of Originating-Terminal and measure the change in prewalkOverlap and offPct · Add an explicit 'unowned-scripts' fallback lane (distinct from any content room) for commits touching paths no room's glob claims, and measure whether the false-positive out-of-lane rate for content rooms (voice, architect) drops over the next 2 weeks

   Measured outcome (deterministic, not self-graded): room mean offPct 32→34 over 6 commits since · dictionary ablation saved 849B

   This event inflates 🎤 voice's apparent loss-ratio with a legitimate ops/analytics build it never actually claimed, so voice's premium should be restored the moment lane-resolution prefers the commit's own glob-computed Relevant-Rooms trailer over Originating-Terminal, rather than by asking voice's future content commits to somehow drift less.

   Open the signed attestation →

6. 2026-09-05 · 📐 VS Code Architect · offPct 25 · structural

   research(tape): three of the four questions already have answers — 823 rows, unread for a month

   Why it left the lane: docs/ops/ is not in 📐 architect's owned-surface glob list at all (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) — so the single new file this commit adds was never going to land inside the room's declared lane no matter what it said. On top of that, the diff's actual vocabulary is saturated with laboratory-domain terms: '.thetacog/lens-sidecar-tape.ndjson', '823 rows', 'LOAD-BEARING', 'retrieved and UNUSED', 'HIT RATE: 50.0%', 'the A/B run of 2026-08-07', 'run-ab.mjs', 'tape-steer.mjs'. The commit message even says it outright: 'laboratory owns the lens and the A/B... architect owns the reader.' Two independent pulls — a directory that belongs to nobody's glob, and content mass that belongs to a different room's reef corpus — both point away from architect's lane in the same direction. No region narrative was cached for this sha, so I can't cite [row,col] coords; the owned-surface-glob mismatch plus the lab-vocabulary density is the positional evidence available.

   The semantic pull: Repeated lab-instrument nouns (lens, tape, rule retrieval, A/B, hit rate) dominate the diff's token mass; there is no docs/strategy- or CLAUDE.md-flavored content anywhere in it to counterweight that pull toward the laboratory's reef corpus.

   Intent–reality gap: The commit message and trailers frame this as an architect deliverable ('architect owns the reader that turns 823 unread rows into a receipt'), filed as an Originating-Terminal: architect commit — but the receipt measured 25% of the lit mass landing outside the claimed lane, and every substantive sentence in the diff is laboratory's lens/tape/A-B analysis, not an architect artifact.

   • git mv the file into docs/ops/laboratory/ (or a laboratory-owned path) so future receipts read its true content lane
   • add docs/ops/laboratory/** to laboratory's owned-surface glob in .workflow/rooms/laboratory.html
   • add docs/ops/architect/** (namespaced, not the flat root) to architect's owned-surface glob in .workflow/rooms/architect.html so architect's own future ops docs have a legitimate home
   • amend CLAUDE.md's COMMIT MESSAGE STRUCTURE section to require docs/ops/<room>/ namespacing and Relevant-Rooms[0] == namespace room
   • backfill: grep docs/ops/ for the other 13 architect out-of-lane commits in the last 14 days and check how many are the same flat-directory pattern before assuming 14 independent drifts

   What we learned: docs/ops/ is an undifferentiated catch-all with no per-room namespace, so a room's measured out-of-lane rate can be partly an artifact of a shared dumping directory rather than 14 independent content drifts — the directory structure itself is uncounted attack surface for the receipt.

   The missing guard (the deliverable): tests/ops/docs-ops-namespaced-by-room.test.mjs — asserts every new docs/ops/*.html file lives under docs/ops/<room-key>/ and that <room-key> is the first entry in that commit's Relevant-Rooms trailer; fails loudly on a flat docs/ops/<date>-*.html add.

   Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized in raw magnitude (138 vs 117 lit, ~85% ratio, no runaway spread) but wrongly directed from the start — this isn't a depth or saturation problem. Adjustment: for docsOnly=true single-new-file commits, inject the committing room's owned-surface glob terms as extra seed weight before the walk runs, so a thin/novel file gets anchored to its claimed room's corpus rather than drifting purely on raw token overlap. Experiments: Re-run this sha's reef seed with architect's owned-surface glob strings (docs/strategy, CLAUDE.md, docs/03-product) injected as seed weight and measure whether prewalkOverlap rises and offPct falls · Recompute the 14-day architect offPct trend with docs/ops/ commits excluded vs included from the baseline, to isolate how much of the room's drift rate is the flat-directory artifact vs real content drift · Build a labeled 'foreign vocabulary' probe: tag known laboratory terms (lens, tape, sidecar, load-bearing, hit rate, retrieval) and measure what fraction of docsOnly commits from non-laboratory rooms spike it, as an early warning ahead of the full walk

   Measured outcome (deterministic, not self-graded): room mean offPct 15→17 over 11 commits since · dictionary ablation saved 1664B

   A docs-only commit whose substantive content belongs to another room's reef territory but ships through an unnamespaced shared directory inflates the filing room's measured drift and understates its actual lane discipline; restoring the premium means namespacing docs/ops/ by owning room so the receipt measures where content actually lives, not where a shared catch-all directory happened to put it.

   Open the signed attestation →

7. 2026-09-05 · 🧪 laboratory · offPct 47 · mistagged

   fix(elicit): the tape decides who said it — a Gemini sentence was wearing the operator's badge, five times

   Why it left the lane: The diff's actual mass sits in three registers the walk reads as B3/Tactics.Signal and C1-C3/Operations.*: (1) a new exported function `attributeQuote({quote, turns})` with four string-literal verdict states ('operator'/'model'/'unattributed'/'unverified') — a SIGNALING/classification vocabulary, not a grid-authoring one; (2) heavy use of 'tape', 'turn_hash', 'thread', 'ladder.ndjson' — data-pipeline/queue nouns that read as Operations.Grid/Loop; (3) the 165-line new test file is built entirely around verifying a SIGNAL (who said it) against a LOG (the tape), which is structurally a Tactics.Signal payload wearing a laboratory/elicit path. The commit message frames this as a `bf-5059` disposal decision (governance/decision correctness) but the diff is 245 lines of signal-attribution plumbing across scripts/elicit + two test files — none of it touches git-discipline surface at all, so the walk correctly finds no C,B3 mass and instead lights up the neighboring Operations/Tactics blocks the vocabulary actually belongs to.

   Intent–reality gap: The commit message claims a laboratory-room decision-ledger fix (bf-5059, bounded to scripts/elicit/elicit-run.mjs) with a governance/authorization framing; the receipt measured the actual reality mass landing across Operations.Grid/Loop/Flow and Tactics.Speed/Deal, i.e. as generic queue/signal-processing work rather than the room-scoped elicitation-ledger fix it was claimed to be.

   • Add a one-line scope note to the commit message template used for bf-* disposal fixes: 'Lane: <room> — <owned surface touched>' so the claimed lane is explicit and checkable against the diff, not inferred from prose.
   • tests/elicit/operator-span-is-verified-against-the-tape.test.mjs: keep it filed under tests/elicit/ (already correct) — do not let a future refactor promote attributeQuote() into a shared lib without re-declaring the lane.
   • Cross-link this commit's `why` block to bf-5059's own ledger entry so the disposal decision and the code fix are traceable as two separate units of work (per DELEGATE-TO-SUBTASKS), since the message currently bundles 'decide the routine's disposal' with 'fix the laundering bug' as one narrative.
   • If attributeQuote() is ever reused outside scripts/elicit/, add its new home to .workflow/rooms/*.html owned-surface globs at that time — do not pre-claim it now.

   What we learned: A commit whose PROSE is framed as a governance/decision-correctness fix (bf-5059 disposal, ratchet authorization) but whose CODE is a narrow signal-classification function will read as generic Operations/Tactics plumbing to the walk regardless of which room authored it — the receipt is grading the diff's vocabulary, not the commit message's stated stakes, and a philosophically load-bearing commit (AXIOM 1, W2, the ratchet) is exactly the kind of message that tempts an outsized lane claim relative to a 3-file, single-function diff.

   The missing guard (the deliverable): A pre-commit check that compares commit-message rhetorical weight (references to AXIOM/HARD RULE/ratchet-language density) against diff size/file-count, warning when a philosophically heavy message wraps a small, single-function diff — not to block it, but to prompt an explicit 'Lane:' scope line before commit.

   Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litReality (132) exceeds litIntent (113) by ~17%, and fill is only 0.98% of the 144x144 space — the walk is shallow and under-saturated, not exploding, but it is still finding enough weak reality-side edges (queue/ledger/tape vocabulary) to fan out into 31 red regions from a single-function diff. Concrete adjustment: lower --max-depth by one ply for elicit/laboratory-room commits specifically, or require a higher gzip-NCD similarity floor before a reality-side cell is allowed to light in a block the intent seed never touched (prewalkOverlap=2 should gate more aggressively than it currently does). Experiments: Experiment: re-run this commit's pipeline with prewalkOverlap gated at a minimum threshold (e.g. require ≥5 before allowing reality-walk expansion past ply 1) and measure whether offPct drops below 15 for this sha without losing true-positive cross-lane detections on the 13 known out-of-lane laboratory commits. · Experiment: seed the reef dictionary with 'attributeQuote', 'operator_quote', 'turns', 'tape', 'ladder.ndjson' as elicit-room-specific vocabulary anchored to B3.Tactics.Signal, then re-render this sha's panel and measure whether litReality concentrates rather than spreading into C1-C3. · Experiment: compare this commit's offPct against a synthetic control where the same diff is committed with a scope-line-annotated message (per spec_rewrite above) and measure whether the walk's seed anchoring changes at all — this tells us whether the drift is message-driven or purely code-vocabulary-driven, which determines whether the fix belongs in commit discipline or in the dictionary.

   Measured outcome (deterministic, not self-graded): room mean offPct —→43 over 4 commits since · dictionary ablation saved 847B

   A 47%-off-lane receipt on a correctness-critical, AXIOM-1-citing fix reads as elevated tail-risk for this lane's loss ratio even though the underlying code change is sound and guarded; premium is restored by the scope-line convention plus the owned-surface glob addition, which together let future elicit/ commits re-anchor to their true lane instead of re-triggering a fresh 31-region spread each time.

   Open the signed attestation →

8. 2026-09-05 · 🔨 builder · offPct 32

   fix(mirror): the push refuses any address outside the allowlist, not just known markers

   Why it left the lane: (sensemaking unavailable — facts-only intervention)

   Measured outcome (deterministic, not self-graded): room mean offPct 24→21 over 6 commits since

   Open the signed attestation →

9. 2026-09-05 · 📐 VS Code Architect · offPct 33 · mistagged

   req(R7): the competence-shape match is unbuilt — and two things in the repo are homonyms for it

   Why it left the lane: The diff is a data-file-only change to data/requirements/2026-09-05-session-asks.json (7 insertions, 18 deletions) — pure requirements-ledger bookkeeping. The commit message, by contrast, is a dense prose diagnosis naming file paths (scripts/pmu/shape-match-ab.mjs, scripts/mesh/mesh-prove.mjs), architectural constructs (lens placement, competence receipt, room-to-room transactions), and a cross-reference to CLAUDE.md's homonym register. The reef's gzip-NCD seed matched on that prose — 'homonym', 'shape match', 'competence', 'marketplace claim' — which reads as Tactics/Operations vocabulary (Speed, Deal, Signal, Grid, Loop, Flow: all words about mechanism and exchange), not as the confidence-bookkeeping lane a requirements-JSON edit should occupy. The walk lit 138 reality cells against 127 intent cells with only 5 prewalkOverlap — the commit's OWN WORDS, not its diff, are what dragged mass into 28 out-of-lane Tactics×Operations coordinates.

   The semantic pull: Terms 'shape-match-ab.mjs', 'mesh-prove.mjs', 'priced competence receipt', 'competence transaction between two room-nodes', and 'O(1) and search-free' are code/architecture nouns describing OTHER files' mechanisms — they gzip-compress near Tactics.Deal/Tactics.Speed/Operations.Grid vocabulary because that's where those two named scripts actually live and are indexed, even though this commit touches neither file.

   Intent–reality gap: The commit message reads like an architecture/code-audit finding (it names and characterizes two implementation files in detail) but the receipt shows the actual reality-lane work was a 7-line trim of a JSON confidence ledger — the message's vocabulary claims a lane the diff never enters.

   • Add a header comment to data/requirements/2026-09-05-session-asks.json documenting the one-sentence-reason convention for confidence-delta commits
   • File the homonym finding (shape-match-ab.mjs vs mesh-prove.mjs vs the marketplace claim) as its own doc under docs/architecture/ or as a CLAUDE.md register line, split out of the ledger commit
   • Add a CLAUDE.md pointer under the PMU homonym section cross-referencing this specific pair (lens-placement shape-match vs room-to-room competence receipt) so a future grep for 'competence shape' finds the disambiguation directly
   • Consider a commit-message length/file-name-density heuristic for pure data-file commits

   What we learned: A commit whose diff is ONLY a requirements-ledger trim can still carry an entire architecture audit in its prose, and the reef has no way to distinguish 'this message documents the diff' from 'this message documents unrelated files it merely investigated' — the message itself is the drift vector even when the diff is inert.

   The missing guard (the deliverable): A pre-commit check for commits touching ONLY data/requirements/*.json: warn (never block) if the commit message body contains ≥2 distinct file-path references (glob-matched against tracked repo files) that are NOT the changed path — this is the structural signature of an audit-report message riding on a data-only commit.

   Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Right-sized on spread (127 vs 138 lit is a modest 8.7% asymmetry) but arguably over-triggered on FILL — 28 red regions off a 2-ply walk from a data-only commit suggests the walk is faithfully propagating a bad seed rather than a bad walk; adjustment: no depth change needed, but consider down-weighting seed terms that are themselves OTHER files' identifiers (mesh-prove.mjs, shape-match-ab.mjs) when the current diff doesn't touch them. Experiments: Run this exact commit through the reef with the commit message truncated to just its first paragraph (drop the two bullet examples naming other files) and measure whether offPct drops below 15 — isolates whether the file-name mentions are the actual drift driver · Re-run prewalkOverlap computation restricted to only the diff's changed lines (not the full commit message) for docsOnly=true commits and compare offPct — tests whether gating the reef seed to the diff instead of the message closes this class outright · A/B the walk at max-depth=1 vs the current 2-ply on the last 10 offPct>30 architect-room commits and measure whether shallowing recovers lane accuracy without losing legitimate cross-file findings

   Measured outcome (deterministic, not self-graded): room mean offPct 23→32 over 3 commits since · dictionary ablation saved 880B

   An out-of-lane architect-room event driven by message-prose rather than diff content is a low-severity, high-frequency false-positive class (34 events/14 days already) that inflates this lane's apparent loss ratio without reflecting real code drift — restoring premium means separating audit-findings prose from data-only commits so the receipt measures actual reality-lane risk, not commentary.

   Open the signed attestation →

10. 2026-09-05 · 🔒 vault · offPct 24 · mistagged

    edit(book): ch12 — the comp is only brilliant if the reader can still see the money

    Why it left the lane: Every changed path in this commit (books/tesseract/chapters-html/12-the-budget-is-the-proof.html, books/tesseract/chapters/chapter-12-the-budget-is-the-proof.md, and the mirrored public/book/chapters/ file) sits entirely outside vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). Nothing in the diff touches a patent, legal, or data-room file. But the prose itself is saturated with patent/insurance/legal vocabulary — 'royalty', 'licence', 'substrate-attested receipt', 'fiduciary policy', 'indemnification', 'VaR', 'reinsurance', 'US 19/637,714, 36 claims, Track One', 'role continuity' — which is exactly the lexicon the vault reef anchors on. So the reef's gzip-NCD seed pulls reality mass onto vault-adjacent coordinates on CONTENT alone, while the commit's real location (book chapter files) has zero overlap with vault's glob. The commit's own Relevant-Rooms trailer already names 🎩 operator, 📐 architect, ☕ network — not vault — yet the event header still reports 'room 🔒 vault', meaning whatever step assigns the scored room is reading Originating-Terminal, not Relevant-Rooms.

    The semantic pull: Legal/insurance-shaped nouns in a book chapter ('fiduciary policy', 'indemnification', 'reinsurance', 'US 19/637,714', 'licence') gzip-match the vault reef's patent/legal vocabulary bank even though the changed files never touch docs/01-business/patents/ or docs/legal/.

    Intent–reality gap: The commit message claims a pure book-editing task (de-jargoning three paragraphs of ch12, fixing a scale-bridge with the Visa/1958 history, fixing the 1985 inevitability claim) with Relevant-Rooms correctly naming operator/architect/network — but the receipt is being scored against vault, a room whose owned-surface glob matches zero of the changed files, so the 'out-of-lane' verdict is measuring the wrong lane rather than a real deviation.

    • Locate the exact function in scripts/pmu/pipeline.mjs (or wherever offPct/room is assigned for this receipt) that reads the commit trailers, and confirm it is keying off Originating-Terminal rather than Relevant-Rooms
    • Add a regression test asserting: given a commit whose Relevant-Rooms excludes Originating-Terminal AND Originating-Terminal's owned-surface glob matches zero changed files, the receipt's scored room must not equal Originating-Terminal
    • Extend the existing e94a8cd14 fix/guard to also cover this asymmetric case (Relevant-Rooms already correct, but the scorer still bypasses it) — the prior fix patched trailer GENERATION, not trailer CONSUMPTION
    • Add a 'zero-owned-surface-match' flag to the per-commit receipt row so future intervention passes can distinguish mistagged-room drift from genuine cross-lane work at a glance
    • Audit the vault room's last 21 out-of-lane events for the same file-path/owned-surface mismatch before attributing the room's rising offPct trend to real drift

    What we learned: Two prior fixes (e6c2aa4ce, e94a8cd14) both patched how the Relevant-Rooms TRAILER gets generated — yet this event shows a correctly-generated trailer (Relevant-Rooms explicitly excludes vault) being ignored anyway, because the SCORING step that assigns 'room' for offPct purposes reads Originating-Terminal directly and was never updated to consult Relevant-Rooms at all. Fixing the trailer generator does not fix the pipeline if the consumer of that trailer never reads the field the fix corrected — the bug moved from generation to consumption, and the room mean offPct kept climbing (11→24, 17→27, 21→12-but-not-zero, now 24 again) because every 'fix' patched a different layer than the one still doing the scoring.

    The missing guard (the deliverable): A test that runs the drift-receipt pipeline over a synthetic commit whose Originating-Terminal is room X, whose Relevant-Rooms excludes X, and whose changed files match zero of X's owned-surface globs — and asserts the emitted receipt's scored room is never X. Verified in both directions (X correctly excluded when the fix is present; X wrongly included on today's unfixed pipeline) per the GUARD-SEEN-TO-FAIL discipline, using this exact commit (9698c342a9) as the real defect fixture.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow — litIntent/walksIntent and litReality/walksReality are both ~100%, meaning every walk step successfully lights a cell with no dead ends to indicate the walk is starved for depth. The adjustment needed is upstream of the walk: narrow the INTENT corpus fence for docs-only commits (docsOnly=true here) so intent's own walk (131) isn't nearly as broad as reality's (133) for what the commit message frames as a scoped 3-paragraph prose fix. Experiments: Re-run this exact receipt scoring the room as 'architect' (first Relevant-Rooms entry) instead of 'vault' and measure whether offPct drops toward architect's baseline — confirms the fix targets the right lever before shipping code · A/B seed the intent walk from the commit message's explicit scope ('edit(book): ch12') vs. current full-message ingestion, and measure whether litIntent narrows without litReality changing, which would tighten sigmaDrift for docs-only commits specifically · Ablation-test adding 'substrate-attested receipt' and 'role continuity' as reef terms bucketed under a distinct 'book/insurance-thesis' anchor separate from the 'legal/patent' anchor, and measure whether vault-attributed book commits' offPct drops without moving real patent-commit offPct

    Measured outcome (deterministic, not self-graded): room mean offPct 31→24 over 6 commits since · dictionary ablation saved 949B

    Scoring book-content commits against vault's patent/legal lane manufactures false-positive drift events that inflate vault's apparent loss-ratio and would mis-price that lane's premium; restoring room-for-scoring to an actual owned-surface match (or Relevant-Rooms) removes the false signal and lets vault's premium track only commits that genuinely touch patents/legal/data-room.

    Open the signed attestation →

11. 2026-09-05 · 📐 VS Code Architect · offPct 31

    retract(lens): R20 was wrong — the reef is not seeded with boilerplate, and the filter has worked since June

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 19→33 over 4 commits since

    Open the signed attestation →

12. 2026-09-05 · 📐 VS Code Architect · offPct 20 · mistagged

    fix(pmu): the mass-floor door existed and was wired to nothing — every panel now carries admissibility

    Why it left the lane: The receipt's own telemetry shows prewalkOverlap=2 — before the ballistic walk even ran, the intent-seeded and reality-seeded reef cells barely touched, so the walk started from two almost-disjoint origins even though it then spread to comparable size on each side (walksIntent=133/litIntent=133 vs walksReality=121/litReality=121, both ~saturated). The two changed files — scripts/pmu/pipeline.mjs and tests/pmu/panel-carries-admissibility.test.mjs — sit entirely outside 📐 architect's declared owned-file-surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md); there is zero glob overlap. The commit's own Story trailer says so explicitly: 'laboratory measured the confound and owns the door... architect owns the wire.' The room label came from which terminal app happened to be open (Originating-Terminal: 📐 VS Code Architect), not from which coordinate the diff actually landed on.

    The semantic pull: Real strings from the diff that pull toward operations/instrumentation anchors the architect room owns no files in: `canonicalMassFloor`, `panel-door.mjs`, `packages/thetacog-mcp/scripts/tape/physics.mjs`, `gzipSync`, `record('admissibility'`. None of these terms, nor their containing paths, appear anywhere in architect's owned-surface list — they are pmu/laboratory vocabulary being committed from an architect-labeled terminal. No region narrative was cached for this sha, so this is read off the diff and file paths, not off a cached coordinate list.

    Intent–reality gap: The commit message claims and closes an architect-room requirement ('R18, the second-ranked open requirement, closed... architect owns the wire that was missing') while the receipt measures the actual mass landing on files that appear in zero architect-owned globs and that the commit's own Story trailer assigns to laboratory and operator instead.

    • Add an owned-file-surface ∩ changed-files check to whatever script attributes an out-of-lane event to a room for the '19 out-of-lane events in 14 days' rollup — fall back to Relevant-Rooms[0] on zero overlap.
    • Update CLAUDE.md's COMMIT MESSAGE STRUCTURE section with the Relevant-Rooms-precedence rule above.
    • Batch-audit the other 18 architect out-of-lane events from the last 14 days for the same zero-glob-overlap signature — if most match, this is confirmed systemic, not incidental.
    • Confirm 🧪 laboratory's owned-surface declaration (.workflow/rooms/laboratory.html) explicitly includes scripts/pmu/** and tests/pmu/** so the fallback has somewhere correct to land.

    What we learned: The room bucket used for this drift-rollup is chosen from Originating-Terminal — i.e. which terminal app happened to be open — not from which room's owned-file-surface the diff actually intersects. A room's measured out-of-lane rate can therefore reflect operator terminal habits rather than code quality, and no amount of correctly-scoped future commits from that terminal will converge it, because the misattribution happens upstream of the offPct computation the prior three interventions were trying to move.

    The missing guard (the deliverable): tests/pmu/room-attribution-uses-file-overlap.test.mjs — asserts that for a synthetic commit whose Originating-Terminal owns none of the changed files but whose Relevant-Rooms names a room that does, the room-level drift rollup attributes the event to that Relevant-Rooms entry, not to Originating-Terminal.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent=133/133 and litReality/walksReality=121/121 — every walk step produced a lit cell on both sides, and 133/144≈92%, 121/144≈84% means the walk is close to full-lattice saturation. At that saturation level the walk has little room left to discriminate in-lane from out-of-lane; adjustment: cap --max-depth lower for commits like this one (small diff, two files) so the walk terminates before near-total coverage flattens the signal. Experiments: Re-render this commit's walk with --max-depth capped at ply 2 and compare offPct against the current ply-unbounded run — tests whether near-saturation (133/144, 121/144) is itself diluting the signal. · Reseed the reef using the union of Relevant-Rooms (laboratory+architect+operator) instead of a single inferred room, then recompute prewalkOverlap — tests whether the overlap=2 floor is a room-seeding artifact rather than genuine semantic divergence. · Pre-inline the physics.mjs floor constant/value into the reality corpus ingest text (rather than leaving it as an import reference) and re-render — tests whether the cross-package reference itself is contributing measurable off-lane mass.

    Measured outcome (deterministic, not self-graded): room mean offPct 18→35 over 4 commits since · dictionary ablation saved 1155B

    A 20%-vs-15% overage on a commit whose own trailers already name the correct owners is a low-severity, self-adjudicated claim, but 19 such events in 14 days is inflating architect's apparent loss-ratio with pure room-attribution noise rather than real drift; premium is restored by fixing the routing rule above, not by tightening review on architect's terminal.

    Open the signed attestation →

13. 2026-09-05 · 📐 VS Code Architect · offPct 28 · mistagged

    fix(R2): stop resuming — carry the history, so the lens fires on the turns it exists to govern

    Why it left the lane: The changed files — scripts/lab/run-turns.mjs and tests/lab/run-turns.test.mjs — carry terms like `runEpisode`, `correctiveTurn`, `carryHistory`, `UserPromptSubmit`, `session` that match none of 📐 architect's owned-surface globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). They match 🧪 laboratory's territory instead — the commit's own Relevant-Rooms trailer names laboratory FIRST, and the Story line says outright: 'laboratory owns the experiment that refused and thereby found this.' The reef placed the mass where the content actually is; the room label attached to the event did not follow it.

    The semantic pull: litIntent=133/litReality=135 with only prewalkOverlap=8 shows the two corpora (message text vs. code text) barely share seed cells before the walk even runs — the message is saturated with lab-experiment vocabulary (episode/turn/session/hook) that has zero gzip-NCD mass against architect's owned docs/strategy or CLAUDE.md prose, so the walk had nowhere in-lane to land.

    Intent–reality gap: The commit's Originating-Terminal trailer books this event to 📐 architect, but both the claimed intent (the commit message, entirely about lab-runner hook behavior) and the measured reality (100% of changed files under scripts/lab/ and tests/lab/) sit in laboratory's territory — the commit itself says so in its own Story trailer, yet the receipt was still scored against architect's lane.

    • Locate the intervention-dispatch script that assigns this drift event to a room (search for where 'this room has N out-of-lane events' is computed) and have it prefer owned-surface glob match over changed files, falling back to Relevant-Rooms trailer, before falling back to Originating-Terminal.
    • Confirm 🧪 laboratory's `.workflow/rooms/*.html` owned-surface declaration explicitly lists `scripts/lab/**` and `tests/lab/**` — add if absent, so glob-matching is unambiguous for future events.
    • Re-audit the room's other 23 out-of-lane events in the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms mismatch pattern before treating the 24-event count as genuine architect scope creep.
    • Add the CLAUDE.md clarification above in the same commit as the dispatcher fix so the rule and the enforcement land together (per CODE DEFINITION-OF-DONE).
    • Add the guard named in missing_guard.

    What we learned: CLAUDE.md already documents, in its own words, that Originating-Terminal and content ownership are permitted to diverge — but nothing downstream reads that permission: the drift-receipt pipeline appears to book accountability to Originating-Terminal, converting a documented, sanctioned convention into manufactured out-of-lane events. This room's 24-events-in-14-days figure may be substantially inflated by commits authored from the architect terminal about laboratory (or other rooms') work, not by real scope creep — the prior three lessons for this room all treated offPct trend as a behavior signal; this event shows the trend itself may be measuring the wrong variable for some fraction of its inputs.

    The missing guard (the deliverable): A test that takes a real commit whose Relevant-Rooms trailer names a room OTHER than Originating-Terminal and whose changed files 100% match that other room's owned-surface glob (fixture: a91298ca5 itself — files under scripts/lab/+tests/lab/, Originating-Terminal=architect, Relevant-Rooms first entry=laboratory) and asserts the drift-dispatcher attributes the event to the glob-matched room, not Originating-Terminal — SMOKE: current logic (if it keys on Originating-Terminal) fails this; NO SMOKE: a commit with matching Originating-Terminal and owned-surface files must still attribute normally.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent/walksIntent ≈ 100% and litReality/walksReality ≈ 100% — every walked cell is lighting up, which is saturation, not a shallow or under-used walk. A saturated walk loses contrast between in-lane and out-of-lane cells because nearly everything it touches lights; tighten the ballistic hit threshold (raise the bar for what counts as 'lit') or shorten --max-depth so the walk stops before full saturation, restoring discriminative spread. Experiments: Re-run this commit's reef seed using BOTH architect's and laboratory's owned-surface corpora combined (since Relevant-Rooms names three rooms) and measure whether prewalkOverlap rises above 8 — tests whether single-room seeding is itself the cause of the low overlap. · Ablation-test seeding the intent walk from the full commit message including trailers vs. stripping trailers, and measure whether including 'Relevant-Rooms: 🧪 laboratory' text pulls sigmaDrift down / offPct below the 15 threshold for this exact commit. · Recompute offPct for this event and the other 23 in-window events using owned-surface glob-matched changed-files as the lane definition instead of text-only gzip-NCD placement, and diff the resulting event counts against the current 24 — a measurable count of how many were misattributed.

    Measured outcome (deterministic, not self-graded): room mean offPct 20→27 over 11 commits since · dictionary ablation saved 964B

    This event should not raise 📐 architect's loss-ratio — the content it measures belongs to laboratory by the commit's own admission — so the lane's premium is restored by re-scoring the room's 14-day event count net of Originating-Terminal-vs-owned-surface mismatches rather than by treating this or similar events as real scope creep.

    Open the signed attestation →

14. 2026-09-05 · 📐 VS Code Architect · offPct 39 · structural

    req(lens): R3 answered — hooks-off is an OVERRIDE failure — and the reef is seeded with boilerplate

    Why it left the lane: The diff (data/requirements/2026-09-05-session-asks.json) does three things: appends R3 and R5 to `closed`, deletes R3 and R5 from `open`, and appends a new R20 entry to `open`. The commit message narrates R3 ("hooks-off is an OVERRIDE failure") and R20 ("reef seeded with boilerplate") at length, but never once mentions R5 — whose closed-array text is "R5 — tape reader committed as scripts/lab/tape-steer.mjs with a guard: never computes the template-fixed ratio, refuses to print a rate below 100 evaluable rows, and states its join weakness rather than dressing it as a finding". That string's vocabulary (tape-steer.mjs, template-fixed ratio, evaluable rows) has zero lexical overlap with the message's lens/reef/hooks-off/seedRules vocabulary — it is real reality mass with no intent anchor, which is exactly the shape prewalkOverlap=4 (out of walksIntent=122) is reporting: almost nothing in the reef seed matched before the walk even ran, because a third of the diff's semantic content was never fed into the intent side at all.

    The semantic pull: R5's closed-array string ('tape reader … scripts/lab/tape-steer.mjs … template-fixed ratio … evaluable rows') is the pull — it is real diff content sitting in an instrumentation-reliability lane distinct from the message's lens/reef lane, and it dragged litReality to 135 while litIntent stayed at 122 with a near-zero prewalk overlap (4).

    Intent–reality gap: The commit claims (in prose) to be about answering R3 and seeding R20 into the register; the diff also silently closes R5 (owner: architect, about scripts/lab/tape-steer.mjs) with no mention anywhere in the message — reality closed three requirements, intent narrated two.

    • Add a `Requirements-Closed: R3, R5, R20` (or equivalent) trailer convention for any commit touching data/requirements/*.json, listing every id added to closed[] or removed from open[]
    • Write scripts/ops/requirements-commit-check.mjs: diff the JSON's closed/open arrays against HEAD~1, extract every touched id, assert each id-token appears somewhere in the commit message; warn-only per the existing 'format violation does NOT block commit' convention
    • Add data/requirements/*.json to a room's owned-surface glob in .workflow/rooms/*.html — currently no room's glob lists it at all, despite R5's own owner field reading 'architect'
    • Backfill-audit recent data/requirements/*.json commits for the same silent multi-id bundling to size the historical count of this exact class

    What we learned: The requirements ledger's own `_contract.sufficientFor` says it is sufficient for 'which asks were made, which are closed' — but the commit convention around it has no mechanism forcing the MESSAGE to name every id the diff closes, so a commit can legitimately and silently close a requirement from a different owner than the one the prose narrates. This is a new, named mechanism (silent cross-owner batch-closure in a shared single-file ledger), not a ranking or reef-tuning artifact — distinct from the prior facts-only lessons for this room, which reported only offPct deltas with no cause.

    The missing guard (the deliverable): scripts/ops/requirements-commit-check.mjs — for any commit touching data/requirements/*.json, extract every id added to closed[]/removed from open[] and fail/warn if that id-token is absent from the commit message body.

    Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. Right-sized, not shallow: litReality > litIntent by a wide margin is the walk correctly surfacing a real vocabulary region (R5's) that the intent seed never touched, not over-saturation. Adjustment: seed the intent side directly from the diff's structured content (every touched JSON `id`/`req`/`why` string) in addition to commit-message prose, so requirement-ledger commits don't rely on the message alone to anchor the walk. Experiments: For commits touching data/requirements/*.json, auto-augment the intent corpus with the literal id/req/why strings of every touched entry (not just message prose), then re-render this commit's receipt and measure whether offPct drops below 15 · Backfill the requirements-commit-check guard over the last 30 days of data/requirements commits and count how many historical offPct>15 events in this room match the unmentioned-closed-id signature, to size the class · Add bare requirement-id tokens (R\d+) to the reef seed allowlist so they seed regardless of surrounding prose density, then re-measure prewalkOverlap on this and adjacent ledger commits

    Measured outcome (deterministic, not self-graded): room mean offPct 19→24 over 6 commits since · dictionary ablation saved 960B

    A silently-bundled cross-owner ledger closure is a traceability gap rather than a code defect, so it nudges this lane's loss-ratio only mildly toward mistagged-intent noise; the id-trailer guard plus explicit joint ownership of data/requirements/*.json restores it.

    Open the signed attestation →

15. 2026-09-05 · 🔨 builder · offPct 21 · structural

    fix(floater): closed it is a one-line pill — it was covering the headline it advertises

    Why it left the lane: The diff is one component file that carries two different kinds of mass at once: pure CSS/layout state (`position: 'fixed'`, `maxWidth: open ? 380 : 'max-content'`, `whiteSpace: open ? 'normal' : 'nowrap'`, `borderRadius: open ? 12 : 999`, `boxShadow: open ? …`) and inline marketing/pricing copy branching (`angle.headline`, the literal strings "Send a key", "You can't buy your own alibi", "$20/agent-year — and it isn't for you"). The commit message frames the whole thing as a layout bug ("closed it is a one-line pill"), so the intent lane got fenced around UI/CSS-in-components. But roughly half the diff's actual edit mass is a copy/positioning decision — which sentence earns the click, label vs. headline — and that content-strategy work is what the receipt's ballistic walk picked up landing outside the fenced lane (litReality 139 vs litIntent 131, an 8-cell/6% overshoot that tracks almost exactly the size of the 'AND THE CLOSED COPY HAD TO EARN THE CLICK' paragraph relative to the CSS paragraph above it).

    The semantic pull: The pull terms are the ones with no CSS meaning at all sitting inside style-adjacent code: 'angle.headline', 'Send a key' (a LABEL), 'champion angle', 'You can't buy your own alibi', '$20/agent-year'. These are pricing/positioning vocabulary, not layout vocabulary, and they are physically interleaved with the `borderRadius`/`boxShadow`/`whiteSpace` properties in the same JSX branch — so the walk cannot separate 'how big is the box' from 'what does the box say' because the source file doesn't separate them either.

    Intent–reality gap: The commit message and its own guard comments describe a sizing/collision fix ('closed it is a one-line pill... a fixed 260px block... is not a collapsed state'), but roughly a third of the diff and three of the four new tests are about WHICH SENTENCE displays when closed — a copy/positioning decision, not a layout one. The receipt measured that second concern as mass outside the declared lane.

    • Split any future floater edit into a layout-only commit and a copy-only commit per ITERATION DISCIPLINE — don't bundle 'the box changed shape' with 'the box says something different'
    • Extract 'Send a key' / the price line / angle.headline usage out of the JSX into src/lib/floaterCopy.ts
    • Keep the two new test families (closed-hugs-content vs closed-shows-headline) as separately named SMOKE tests — they already split correctly in this diff, just the commit and message didn't
    • Add a CLAUDE.md example under ITERATION DISCIPLINE naming 'CSS/layout change bundled with copy/positioning change' as a concrete violation class, since the existing text is abstract ('a new section, a renumbering pass') and missed this shape
    • Cross-check whether angle/copy logic already lives in a shared file elsewhere in the repo (the prior test 'renders the ranked angles... setI(n)') before creating a new one — extend it rather than fork a second copy source

    What we learned: This room's prior three out-of-lane events all shipped with 'sensemaking unavailable' and each one's offPct trend got WORSE afterward (16→46, 15→16, null→26) — meaning facts-only intervention with no mechanism diagnosis doesn't reduce recurrence. This event is the first with a real mechanism: a component file in src/components/ that interleaves CSS style objects with inline marketing copy will structurally keep producing multi-domain commits no matter how carefully the message is worded, because the coupling lives in the source, not the prose.

    The missing guard (the deliverable): A pre-commit check on `src/components/**/*.tsx` diffs that flags when a single hunk touches BOTH a `style={{...}}`/CSS-property line AND a string literal containing pricing/copy vocabulary (a small denylist: '$', 'angle.', a quoted sentence >20 chars in JSX text position) — and either requires the commit message to name both concerns explicitly or suggests splitting into two commits.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated: walksIntent(131)==litIntent(131) and walksReality(139)==litReality(139) means every walk step lit a cell — no wasted budget, no truncation flag. No depth adjustment needed here; the lever is seed quality (prewalkOverlap), not walk depth. Experiments: Strip the commit's copy-strategy paragraph ('AND THE CLOSED COPY HAD TO EARN THE CLICK...') from the intent corpus and re-render this same sha's receipt; if offPct drops toward/under 15, the bundled-message language is itself widening the fenced lane beyond the code's actual CSS scope · Re-run prewalkOverlap with pricing/copy tokens ('$20/agent-year', 'angle', 'headline') masked out of the reality-side ingest for this file, and check whether overlap rises above 3 — tests whether copy vocabulary is diluting the seed match to a pure-UI intent lane · After the proposed code_rewrite (copy extracted to floaterCopy.ts), re-render the receipt for an equivalent future layout-only diff and confirm offPct stays under 15 with no copy vocabulary present in the reality corpus at all

    Measured outcome (deterministic, not self-graded): room mean offPct 25→31 over 6 commits since · dictionary ablation saved 897B

    A landing-page CTA drifting 21% out of lane on a revenue-facing component is a modest, low-severity mark against the builder room's loss ratio (6 points over threshold, no truncation, no ingest-health flags) — restored by extracting the copy from the layout so each future commit's claimed lane matches its actual edit mass.

    Open the signed attestation →

16. 2026-09-05 · 🔨 builder · offPct 33 · structural

    feat(iamfim-report): --since <1h|90m|2d>, and an empty window says so in words

    Why it left the lane: The whole diff lives in scripts/track/iamfim-visits.mjs — a data-pipeline/reporting script full of analytics vocabulary ('gradeVisit', 'etParts', 'collapseViews', 'windowNote', 'geolocate', 'iamfim-visits-window.html'). None of that vocabulary or path matches 🔨 builder's declared owned surface (src/app/, src/components/, src/lib/, public/, the two book-build scripts, hooks/pre-commit, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json) — every one of those is UI/component/build tooling, and 'scripts/track/**' is claimed by none of them. The reef has nowhere honest to seed this file's terms inside builder's lane, so the walk's reality side (litReality=126) necessarily spreads past intent (litIntent=111) into whatever adjacent block the analytics vocabulary happens to fall in.

    The semantic pull: Terms like 'iamfim-visits.mjs', 'gradeVisit', 'windowNote', 'Nothing landed in this window.' and the geo/grading machinery already living in the file are operations/reporting diction, not builder's app-surface diction — they pull placement toward an operational-instrumentation region the reef has no bridge back to src/app or src/components.

    Intent–reality gap: The commit message and Originating-Terminal both read as 🔨 builder work, but the entire change is inside scripts/track/iamfim-visits.mjs, a path no room's owned-surface glob (as given to builder) actually covers — the receipt is correctly reporting that reality landed on a path with no legitimate claimant in builder's lane, even though the code itself is coherent and well-scoped to its own (unlabeled) domain.

    • Confirm which room's .workflow/rooms/*.html 'Owned file/dir surface' section should list scripts/track/** and add it there
    • Audit scripts/ for sibling orphaned paths sharing this fate (scripts/self-prompt*, scripts/predict-leverage.mjs, scripts/print-leverage.sh) against all 9 rooms' globs
    • Add 'visits', 'tracking', 'iamfim', 'geolocate' to the correct room's vector_keywords in data/rooms.json so sync-vectors projects this file correctly going forward
    • Pull the room's last 69 out-of-lane events and tag each by orphan-path vs genuine-mistag vs genuine-scope-creep to confirm this is systemic before assuming one glob-add fixes it
    • Write the orphan-path coverage guard (see missing_guard) and run it once over the full repo to surface every currently-unclaimed path, not just this one

    What we learned: This room's chronic high offPct (69 events/14d, prior trend 16→46, 15→16) has never had a sensemaking pass before now — the first one shows the signal isn't catching scope creep at all, it's catching a structural hole in the room-ownership registry: scripts/track/ (and likely other scripts/ subdirectories) is claimed by zero rooms, so ANY terminal that touches it registers 100% out-of-lane regardless of who's actually working correctly.

    The missing guard (the deliverable): A commit-time (or periodic) check that every changed path matches at least one room's declared owned-surface glob, distinguishing 'orphan path' (matched by zero rooms — a registry gap) from 'cross-lane path' (matched by a different room than the committer — a real mistag). Something like tests/rooms/owned-surface-coverage.test.mjs asserting full glob coverage over scripts/, src/, docs/.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Both litIntent/walksIntent and litReality/walksReality sit at 100% — the walk is saturated, not sparse, and walkTruncated=true means it hit the depth budget before naturally terminating. A saturated + truncated walk can't show where it would have settled. Concrete adjustment: re-run this commit's walk with --max-depth doubled and compare offPct; if it drops, truncation (not genuine drift) was inflating the figure. Experiments: Re-run the ballistic walk on c17d7755f with --max-depth doubled; measure whether offPct converges lower once the walk isn't cut off mid-recursion · Build the orphan-path coverage guard and re-score this room's last 69 out-of-lane events, partitioning by orphan-path vs mistag vs scope-creep; measure what fraction the ownership gap alone explains · Seed the reef with a dedicated scripts/track/ analytics-vocabulary block (visits, grade, window, cutoff, geolocate) so prewalkOverlap rises above 4 on the next commit to this directory, regardless of which room's terminal makes it

    Measured outcome (deterministic, not self-graded): room mean offPct 26→30 over 4 commits since · dictionary ablation saved 1256B

    An out-of-lane event caused by a registry gap rather than real scope creep is a false loss signal against this room's lane — leaving it unaddressed inflates 🔨 builder's apparent loss-ratio on work it did correctly, and premium is restored the moment the ownership glob (not the code) is corrected.

    Open the signed attestation →

17. 2026-09-05 · 🔨 builder · offPct 28 · structural

    fix(floater): the card stops where the crew bar starts, measured off the bar itself

    Why it left the lane: The diff's line-mass is majority tests/: GiftLicenseFloater.tsx carries 48 insertions in src/components/ (Builder's own surface), but gift-license-floater.test.mjs (61) + sticker-crew-banner.test.mjs (4) = 65 insertions land in tests/components/ — 60% of the commit's new lines. `.workflow/rooms/cursor-laboratory.html:711` explicitly lists `tests/` as 🧪 Lab Cursor's owned surface ('tests/, audit script logic in scripts/voice-audit-*.sh'), and line 729 of that same file states the house rule directly: 'Hook rule specs in docs/03-product/ are paired with a failing test in tests/ before Builder ships.' `.workflow/rooms/iterm2-builder.html`'s own 'Owned file/dir surface' section never lists tests/ anywhere. So the strings that dragged reality mass out of lane are exactly `querySelector('[data-crew-hook]')`, `getBoundingClientRect().height`, `new ResizeObserver(measure)` and the four 'THE BUG:' assertion blocks — all real code, all correctly written, all physically inside a directory this room's own declared surface excludes.

    The semantic pull: Not a content-semantics pull so much as a directory-ownership one: the walk lit tests/components/gift-license-floater.test.mjs and tests/components/sticker-crew-banner.test.mjs at comparable density to the component file itself (litReality=135 close to litIntent=141), because those test files quote and re-derive the exact same identifiers (TOP, GUTTER, footClear, data-crew-hook, zIndex) that anchor the intent lane — high textual overlap with the claimed lane, wrong directory for the claimed room.

    Intent–reality gap: The commit message claims a single, contained fix ('the card stops where the crew bar starts') scoped to the floater component — a Builder-lane claim. The receipt measured 28% of lit reality mass outside that lane, and the reason is structural: CODE DEFINITION-OF-DONE mandates the guard ship in the SAME commit as the fix, but the repo's own room map assigns tests/ to a different room (Lab Cursor) than the one whose surface (src/components/) the fix lives in.

    • Add the companion-test clause above to .workflow/rooms/iterm2-builder.html's Owned file/dir surface section
    • Add the mirror-image carve-out to .workflow/rooms/cursor-laboratory.html so its tests/ claim explicitly excludes component-adjacent guard tests owned by their source file's room
    • Run /recontext to check whether any OTHER 68 out-of-lane events in this room's 14-day window are the same tests/-vs-src/components/ split before treating this as commit-specific
    • Audit the other 8 rooms' owned-surface globs for the same gap (CODE DEFINITION-OF-DONE applies repo-wide, so every room that ships fixes will hit this)
    • File the gap as a graduation into CLAUDE.md per DELEGATE-TO-SUBTASKS ('a dispatched agent whose coordinate owns no code surface must graduate that rule into CLAUDE.md') since this is a cross-room rule conflict, not a single commit's error

    What we learned: Two HARD RULES collide silently: CODE DEFINITION-OF-DONE requires every fix to ship its guard test in the SAME commit, and the room ownership map assigns tests/ wholesale to Lab Cursor. Every room that correctly follows the guard rule will therefore generate positional drift the moment more than ~15% of its diff is test code — this is not a one-off placement error, it is a standing structural tax the ownership model levies on rule-compliant commits, and it is new information relative to the three prior 'facts-only' events for this room (none of which could name a mechanism).

    The missing guard (the deliverable): A companion-path normalizer in the lane-attribution step: before the receipt scores offPct, any tests/**/<name>.test.mjs whose body does `readFileSync`/`require`/`import` against a source path already inside the committing room's owned surface in the SAME commit is reclassified into that room's lane rather than scored against tests/'s owner. Concretely: tests/components/gift-license-floater.test.mjs and tests/components/sticker-crew-banner.test.mjs both readFileSync src/components/GiftLicenseFloater.tsx (and the latter also src/components/CrewBar.tsx) — that read is the detectable signal.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or runaway: both sides hit 100% lit-to-walked ratio (141/141 intent, 135/135 reality) with only a 4% count difference — the walk terminates cleanly and isn't truncated (walkTruncated=false). No depth adjustment needed here; the signal to chase is upstream at the reef-seed/room-attribution stage, not the walk's --max-depth. Experiments: Strip tests/ from the reality corpus for this commit, rerun the pipeline, and compare offPct — if it drops near/under 15, the companion-test hypothesis is confirmed as the dominant mechanism, not a coincidental correlation · Recompute prewalkOverlap for gift-license-floater.test.mjs alone vs sticker-crew-banner.test.mjs alone (the latter also reads CrewBar.tsx) to see whether the cross-component test file is a disproportionate contributor to the 28% · Implement the companion-path reclassification as a dry-run flag first (log what WOULD move lanes without changing the receipt) over the room's last 68 out-of-lane events, to size the fix before shipping it live

    Measured outcome (deterministic, not self-graded): room mean offPct 25→33 over 4 commits since · dictionary ablation saved 1025B

    An offPct=28 event whose root cause is a documented cross-room ownership gap (not sloppy placement) is a cheap loss to reprice — the premium on this lane restores the moment the companion-test carve-out lands in both room files, because the same mechanism, left unpatched, will keep firing on every future guard-adding commit in any of the 9 rooms.

    Open the signed attestation →

18. 2026-09-05 · 🎭 Performer · offPct 27 · mistagged

    fix(iamfim): the disc was an intermission — the ask moves below the beat that earns it

    Why it left the lane: The commit's Originating-Terminal trailer files this under 🎭 Performer, whose owned surface is exclusively decks/fundraising/event prose (docs/decks/, docs/strategy/deck/, docs/01-business/fundraising/, docs/strategy/event-*). The actual diff is 106 lines of src/app/iamfim-landing/page.tsx (React/TSX markup, className strings, JSX comment blocks) plus a 192-line new test file under tests/surfaces/. Neither path matches any glob in Performer's owned-surface list. The commit's own Relevant-Rooms trailer already names the correct owners — operator, architect, voice — not Performer. So the out-of-lane reading isn't the walk misbehaving; it's the receipt trusting the wrong trailer field (Originating-Terminal, 'which terminal the operator was physically in') as the lane label instead of Relevant-Rooms (the actual owned-surface match), and CLAUDE.md itself documents that these two fields are allowed to diverge.

    The semantic pull: The vocabulary in the diff — 'insurable', 'underwriter', 'If it's debatable, it's not insurable', 'pitch'-shaped aphorisms — sits in the same lexical neighborhood as Performer's actual pitch-deck material (docs/decks/insure-rounds/ is literally about insurability and underwriting), so the gzip-NCD reef seed found a legitimate low-level overlap (prewalkOverlap=5) on subject matter. But the ballistic walk then had to keep expanding to cover where the mass actually lives — real JSX (`data-lane="disc-place-one"`, `className="mt-9 max-w-3xl font-serif..."`) and a new Node test file — and those cells belong to the frontend/test surface, which is nobody's Performer-owned glob. Same words, different file lane.

    Intent–reality gap: The commit message frames this entirely as a reader-experience/copy-ordering fix narrated in prose ('the disc was an intermission', 'the ask moves below the beat that earns it') with no claim about which room's surface it touches — but the receipt attributes it to Performer's lane via Originating-Terminal, and 27% of the reality mass lands on file paths (src/app/iamfim-landing/page.tsx, tests/surfaces/) that only Relevant-Rooms (operator/architect/voice) — not Performer — legitimately owns.

    • Patch CLAUDE.md's COMMIT MESSAGE STRUCTURE section with the Relevant-Rooms-is-authoritative-for-lane clause
    • Patch the receipt's room/lane-resolver to prefer Relevant-Rooms/glob-match over Originating-Terminal
    • Add tests/pmu-simulator/receipt-lane-uses-relevant-rooms.test.mjs fixed on this commit (c72410d69) as the historical defect fixture
    • Audit other 🎭 Performer out-of-lane events for the same Originating-Terminal-vs-Relevant-Rooms mismatch before trusting the room's offPct trend
    • Add an explicit negative note to Performer's owned-surface declaration that src/app/** and tests/** are never in its lane, sharpening future glob matches
    • Run the compression_dictionary_adds ablation before adopting any of the proposed dictionary strings

    What we learned: CLAUDE.md explicitly sanctions Originating-Terminal diverging from a commit's true owned-surface room ('that's fine and worth recording'), but the drift-receipt pipeline was never told which of the two trailer fields is authoritative for lane-membership scoring — so a commit can be perfectly, correctly tagged per the repo's own convention and still trip an out-of-lane receipt purely because the receipt reads the wrong field.

    The missing guard (the deliverable): tests/pmu-simulator/receipt-lane-uses-relevant-rooms.test.mjs — feed it this exact commit (Originating-Terminal=Performer, Relevant-Rooms=operator/architect/voice, changed files=src/app/iamfim-landing/page.tsx + tests/surfaces/iamfim-fold-places-the-reader.test.mjs) and assert the resolved lane is drawn from Relevant-Rooms/owned-surface glob match, never equal to 'performer'.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litIntent/walksIntent = 143/143 and litReality/walksReality = 127/127 — every visited cell counts as 'lit' on both sides, i.e. the walk is saturated rather than discriminating; there is no filtering happening between 'visited' and 'signal'. Concrete adjustment: require an anchor to share BOTH row and column axes with a seed (not either), which should shrink litIntent/litReality below walksIntent/walksReality and make offPct reflect genuine cross-lane mass rather than every cell the walk merely touched. Experiments: Re-run the reef seed splitting doc-prose ingest ('insurable','underwriter') from code-identifier ingest ('data-lane','className') per the PMU DOGFOOD semantic-ingest rule, and measure whether prewalkOverlap rises or offPct falls for this commit · Recompute this and the last N Performer-tagged commits using Relevant-Rooms-as-lane instead of Originating-Terminal-as-lane, and measure the change in the room's offPct trend · Tighten the lit-threshold to require shared-row-AND-column anchoring instead of shared-row-OR-column, re-run over recent history, and measure litIntent/litReality shrinkage against offPct stability

    This event is a false claim against Performer's loss ratio — the room did nothing wrong, so the fix that restores its premium is correcting the receipt's lane-resolver, not asking Performer to change behavior.

    Open the signed attestation →

19. 2026-09-05 · 📐 VS Code Architect · offPct 21 · mistagged

    lab(tape): the join was the weakness, not the lattice — 9 evaluable rows became 126, and it reproduces

    Why it left the lane: Every changed line in the diff sits inside scripts/lab/tape-steer.mjs and tests/lab/tape-steer.test.mjs — the new overlap()/tokens()/bestMatch() containment-Jaccard matcher, the OVERLAP_MIN=0.5 threshold, and the pool-membership rewrite. None of those paths appear in the architect room's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The commit's own trailers already name the true owner — 'Relevant-Rooms: 🧪 laboratory, 📐 architect, 🎩 operator' and 'Story: laboratory owns the join and the split still owed' — so the diff's semantic pull (join/matcher/z-score/threshold vocabulary) is entirely laboratory-shaped content wearing an architect Originating-Terminal stamp.

    The semantic pull: Terms like 'containment overlap', 'Jaccard', 'z = 85.4', 'pool membership uses the same matcher', and file paths scripts/lab/*, tests/lab/* are all measurement/statistics vocabulary that the reef would seat near laboratory-owned coordinates, not near the strategy/spec-doc vocabulary (CLAUDE.md, docs/strategy/) that defines the architect room's actual lane.

    Intent–reality gap: The commit is stamped 'Originating-Terminal: 📐 VS Code Architect' and the receipt scores it against the architect lane, but 100% of the changed files and 100% of the commit's own narrative ('laboratory owns the join... architect owns the tool, which now answers instead of refusing') describe laboratory-owned statistical tooling — the claimed room and the measured content disagree from the commit's own text, not just from the receipt.

    • Patch CLAUDE.md § COMMIT MESSAGE STRUCTURE with the glob-precedence rule above (spec_rewrite text).
    • Add tests/pmu-simulator/room-attribution-matches-file-glob.test.mjs asserting a commit like ca2af46c6 (100% of diff paths under one room's owned globs, a different Originating-Terminal) is reclassified to the glob-owning room for lane purposes.
    • Confirm .workflow/rooms/laboratory.html's owned-surface glob list explicitly includes scripts/lab/** and tests/lab/**; add if missing.
    • Audit the room's other 31 out-of-lane events from the last 14 days for the same split-ownership signature (Relevant-Rooms names a different single room than Originating-Terminal, with 100% file-glob match) before treating each as a distinct root cause.
    • Add a field to the receipt row (alongside docsOnly/ingestSuspect) flagging 'terminal-glob-mismatch: true' when 0% of a commit's changed paths fall under Originating-Terminal's owned surface.

    What we learned: This room's three prior lessons all measured genuine content-drift interventions with no attribution check — this event shows at least some of the 32 architect out-of-lane events in 14 days may be measuring a labeling artifact (Originating-Terminal vs. actual file-glob owner) rather than the architect room actually doing lab-flavored work; the fix is a routing correction in the receipt pipeline, not a content fix in the commit.

    The missing guard (the deliverable): A guard asserting that when a commit's Relevant-Rooms trailer names one room and that room's owned-surface globs cover 100% of the diff's changed files, the receipt's lane comparison must use that room — never the Originating-Terminal alone — so a physically-present-but-not-actually-working room can't accrue false out-of-lane events on someone else's tooling commit.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated: walkTruncated=false and litReality/walksReality==134/134, litIntent/walksIntent==121/121 means every walk that started also lit, with no budget ceiling hit. The 134-vs-121 reality-wider gap likely reflects the genuinely broader semantic footprint of statistics/measurement code versus a narrower doc-claim, not misconfiguration. Concrete adjustment: re-run with --max-depth+1 on this commit only and check whether litReality grows past 134 — if it plateaus at ~134 the walk is already saturated at the correct depth; if it keeps growing, the depth budget is under-set for tooling commits. Experiments: Re-seed the reef using the Relevant-Rooms trailer text in addition to the commit message body for this commit and 10 similar split-ownership commits; measure whether prewalkOverlap rises above 10 when the disclosed room-split is included as seed text. · Recompute offPct for ca2af46c6 and its 3 prior-lesson siblings using file-glob-owner-as-room instead of Originating-Terminal-as-room; measure whether the room's 14-day mean offPct (currently tracked at 59→46, 59→23, 45→47 in prior lessons) drops when attribution is glob-corrected. · Add the proposed terminal-glob-mismatch flag to the last 90 days of receipts and measure what fraction of all out-of-lane events across all 9 rooms carry it — this decides whether the fix is architect-specific or systemic.

    Measured outcome (deterministic, not self-graded): room mean offPct 23→26 over 4 commits since · dictionary ablation saved 1151B

    An out-of-lane event caused by attribution routing rather than actual scope creep inflates this room's measured loss-ratio with noise, which is restored not by the room doing less lab-flavored work but by the receipt pipeline reading the commit's own disclosed Relevant-Rooms split before pricing the drift.

    Open the signed attestation →

20. 2026-09-05 · 🔨 builder · offPct 18

    fix(iamfim-report): the operator is not a visitor — mark the network, never the city

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 26→30 over 4 commits since

    Open the signed attestation →

21. 2026-09-05 · 📐 VS Code Architect · offPct 23 · structural

    fix(mirror): restore the exec bit the awk rewrite dropped

    Why it left the lane: The diff carries zero text content — it is a pure `old mode 100644`/`new mode 100755` git metadata line with no code tokens, no identifiers, no comments. The gzip-NCD reef seed has nothing semantic to grip on scripts/mirror-to-public.sh's actual content (mirror, publish, sync logic), so the walk fell back to whatever thin signal the commit MESSAGE supplied ('awk', 'rewrite', 'exec bit', 'mv', 'script', 'runnable') — vocabulary that reads as generic tooling/process language and drags placement toward the B-row Tactics coords (Speed/Deal/Signal) and C-row Operations coords (Grid/Loop/Flow) rather than anywhere near the file's real git-discipline lane. With a 0-byte diff body, the walk has no reality mass to anchor on at all, so it spreads across whatever lit cells the message's process-y words touch.

    The semantic pull: Terms 'awk-based insertion', 'mv', 'restore the exec bit', 'directly runnable' in the commit message pulled toward B1.Tactics.Speed (throughput/velocity framing of 'runnable') and C1-C3.Operations.Grid/Loop/Flow (mechanical process framing of a file-rewrite pipeline) — the encircled reds are exactly the Tactics×Operations cross-product, meaning the walk read this as a PIPELINE-MECHANICS commit rather than a one-line permission-bit fix.

    Intent–reality gap: The commit message claims a narrow, mechanical one-line fix (restore chmod +x on a script) — genuinely true and genuinely small. But the receipt's walk, starved of real diff content (a mode-only change produces no textual reality corpus), filled the vacuum with the message's own process vocabulary and mapped the commit across 31 out-of-lane Tactics/Operations cells it never touched.

    • Confirm the reality-corpus ingest for mode-only diffs (scripts/pmu/pipeline.mjs ingest stage) falls back to the commit message ONLY when the diff body is empty, and flag that fallback explicitly in the receipt (e.g. `ingestSuspect:true`) rather than leaving `ingestSuspect=false` on a 0-byte diff.
    • Add a dedicated 'metadata-only commit' detector (old mode/new mode, rename-only, symlink-only) that short-circuits the walk to a direction-only 0-region receipt instead of running the full ballistic walk against the commit message alone.
    • Verify docsOnly/ingestSuspect flag semantics cover the permission-bit-change case — right now this receipt reports docsOnly=false, ingestSuspect=false for a commit that is neither docs nor semantically live code.
    • Cross-check the last 14 days of 📐 architect out-of-lane events for the same 'zero-diff, message-only' signature — if several of the 36 events share this shape, the fix belongs in the ingest stage, not per-commit.

    What we learned: A commit whose diff is pure git metadata (mode/rename/symlink change, zero content bytes) gives the reality-corpus ingest nothing to walk on, so the receipt silently substitutes the commit MESSAGE as the reality corpus — turning every metadata-only commit into a message-vs-message comparison that can drift anywhere the message's vocabulary happens to touch. This is a distinct ingest-health gap from the three prior architect lessons (which were facts-only, no mechanism); it names WHY zero-diff commits are structurally unmeasurable by the current walk.

    The missing guard (the deliverable): A test asserting that for any commit where the diff body is empty (only mode/rename headers, no +/- content lines), the pipeline emits ingestSuspect=true and skips the full walk in favor of a direction-only receipt — never runs sigmaDrift/offPct math derived solely from commit-message vocabulary.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. litReality (101) exceeds litIntent (79) by a wide margin with walksReality > walksIntent too — the walk is over-saturating on the reality side relative to a near-empty diff, meaning it's expanding into unrelated cells rather than terminating early on thin content. Adjustment: cap walk depth/fill proportional to actual diff byte-count, so a 0-byte-content commit gets ply<=1 instead of the full 2-ply spread seen here. Experiments: Re-run this exact commit (d4c260aeac) through the pipeline with the diff stat forced to its true empty-content state and measure whether offPct drops when ingestSuspect gating is added — expect offPct to fall toward the direction-only floor. · Ablation: strip the commit message's process nouns ('awk', 'rewrite', 'mv', 'script') from the reef seed for this commit and re-measure offPct, to confirm the message vocabulary (not the diff) is the actual drift source. · Add a diff-byte-count column to the receipt row and correlate it against offPct across the last 120 commits (n120 series) to quantify how much of the room's 60% breach rate is explained by low-content commits.

    Measured outcome (deterministic, not self-graded): room mean offPct —→12 over 3 commits since

    A zero-content commit inflating offPct to 23 on a lane with a 60% breach rate over 120 commits erodes confidence in the receipt as an underwriting signal for this room; restoring it requires the ingestSuspect gate so metadata-only commits stop counting as loss events at all.

    Open the signed attestation →

22. 2026-09-05 · 🔒 vault · offPct 40

    edit(book): the jargon is a deleted subject, not a hard word — R08a and six repairs

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 34→16 over 4 commits since

    Open the signed attestation →

23. 2026-09-05 · 📐 VS Code Architect · offPct 21 · structural

    rule(ops): a deletion is not a fix, and a guard written in the same breath freezes the judgment

    Why it left the lane: The commit's message is entirely governance/judgment prose — 'framing', 'FORECLOSES', 'deletion is not a fix', 'a guard carries authority a comment does not' — the vocabulary CLAUDE.md rule-writing has always gzip-matched into architect's A.Strategy lane. But 218 of 229 inserted lines (95%) are not that prose: scripts/ops/guard-freezes-judgment.mjs (129 lines of execFileSync/git-diff/regex code) and tests/ops/guard-freezes-judgment.test.mjs + tests/surfaces/iamfim-claims-survive-diligence.test.mjs (89 lines of assert.ok/test() code) are concrete tooling-implementation text that reef-seeds into an Operations/tooling coordinate, not a Strategy one. The room's owned-surface glob (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, CLAUDE.md) never claims scripts/ops/ or tests/ops or tests/surfaces/ at all, so the walk correctly measured the mass where it landed — it just landed somewhere the room's declared lane doesn't cover.

    The semantic pull: The literal strings pulling the walk off-lane are the code identifiers themselves: `execFileSync`, `git diff --name-status`, `assert.ok(!`, `BAN = /assert\.ok\(/`, `changedFiles()`, `deletedCopy()` — dense, repeated implementation vocabulary concentrated in scripts/ops/guard-freezes-judgment.mjs and the two new/edited test files, none of which sit under the room's declared globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md).

    Intent–reality gap: The commit frames itself as architect governance work ('architect owns the guard suite this exposes a hole in') and the only file it explicitly claims via the room's owned surface is CLAUDE.md — but the receipt measured 95% of the actual inserted mass (script + two test files) sitting outside every glob that surface declares, so intent claimed a governance lane while reality mostly executed in a tooling lane.

    • Edit .workflow/rooms/architect.html (or wherever the room's owned-surface list is declared) to add the conditional scripts/ops + tests/{ops,surfaces} clause from spec_rewrite.
    • Add one line under 'A DELETION IS NOT A FIX' in CLAUDE.md pointing at its own guard-freezes-judgment.mjs + its two test files as the in-lane exception this rule creates for architect — the rule's own header should declare this the way it already declares FORECLOSES:.
    • Pull the last 13 out-of-lane events logged for this room and check how many share this same 'CLAUDE.md rule + scripts/ops + tests/ops or tests/surfaces' file signature before committing further to the glob widening — confirm the hypothesis generalizes rather than fitting only this commit.
    • Add the missing_guard test below in the same PR that edits the owned-surface file, so the fix and its own regression check land together.
    • Note in rules-for.mjs's retrieval index (the gap this same commit already names) that a CLAUDE.md HARD RULE block and its cited guard/test paths should be treated as one coordinate cluster, not scored independently by path prefix.

    What we learned: This repo's own ratchet — 'add the guard in the SAME commit as the fix', 'EVERY GUARD SHIPS WITH A FIRING TEST' — structurally guarantees that any room authoring a new CLAUDE.md HARD RULE will also produce script+test mass outside that room's declared owned-surface glob, because guards live in scripts/ops/ and tests/{ops,surfaces}/ by convention and no room's glob claims those paths conditionally. The lane declaration and the commit-bundling discipline are in tension with each other, not with the operator's judgment — which is why 13 of this room's last-14-days events went out of lane and the two prior facts-only interventions moved offPct in opposite directions (46 then 23 then back up to 47): none of them touched the actual mismatch, because none had the sensemaking to see it was structural rather than a sequence of individually bad calls.

    The missing guard (the deliverable): A guard that, for every new HARD RULE block added to CLAUDE.md in a commit, resolves the guard/test file paths it cites ('Guarded by ...') and asserts each path's directory is present in the authoring room's owned-surface glob (from the commit's Originating-Terminal trailer) OR that the rule's own header carries an explicit lane-widening declaration — so a rule that ships its mandatory guard into an unclaimed directory is flagged at commit time instead of surfacing 13 commits later as unexplained room drift.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated: walksIntent and walksReality are within ~11% of each other and neither side hit its depth ceiling (walkTruncated=false). No depth adjustment indicated. One concrete addition: log the litIntent/litReality ratio alongside offPct per commit so a future owned-surface glob change can be checked against whether the RATIO moves (it shouldn't, since the glob only affects attribution) versus whether offPct moves (it might, if the message is also rewritten to name the tooling work explicitly) — conflating the two would misattribute a labeling fix as a walk fix. Experiments: Re-seed this commit's reef placement using the diff content with 'Guarded by', 'ENFORCEMENT:', and other repeated CLAUDE.md boilerplate phrases stripped from both intent and reality corpora, then re-measure prewalkOverlap — if it rises materially, boilerplate density (not real semantic divergence) is inflating this offPct. · Ablation-test adding 'FORECLOSES:' and 'guard-freezes-judgment' to the reef dictionary, then rerun this and the next 5 CLAUDE.md-rule-plus-guard commits and compare offPct before/after — measurable outcome: does the class of commit converge toward a single stable coordinate instead of scattering. · Test splitting ingest for any commit that both edits CLAUDE.md and adds a scripts/ops/tests/ops file into two separate corpora (prose-rule stream, script-implementation stream), score each against its own best-fit lane, and compare the max of the two offPcts against today's single-stream offPct — measurable outcome: whether dual-scoring would have kept this commit under the 15 threshold on its stronger-fitting half.

    Measured outcome (deterministic, not self-graded): room mean offPct 19→10 over 10 commits since · dictionary ablation saved 1111B

    Thirteen out-of-lane events in 14 days for one room reads, on a naive premium model, as a systematically leaky lane — but the mechanism here is a declared, structural glob gap rather than accumulating bad judgment, so the correct restorative is the one-line owned-surface widening plus its guard, not a punitive premium hike on the architect lane.

    Open the signed attestation →

24. 2026-09-05 · 📐 VS Code Architect · offPct 27 · structural

    feat(ops): the envelope gets its first quantitative rung — and the incident that needed it is mine

    Why it left the lane: The entire diff (scripts/ops/spend-cap.mjs, tests/ops/spend-cap.test.mjs) lands in a path no room owns and architect certainly doesn't — architect's owned surface is docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md. Zero of the changed bytes touch CLAUDE.md. What pulled the reef's placement toward architect anyway is the PROSE REGISTER: the commit message is written in CLAUDE.md's own rule-authoring dialect — bold ALL-CAPS declarative headers ('TWO AXES, NEVER MERGED.', 'EVERY OVERRIDE CARRIES A TTL AND THE TOOL REFUSES ONE WITHOUT.', 'OVER-CAP EXITS 75, NOT 1.'), a 'Guarded by tests/ops/spend-cap.test.mjs' citation, and an explicit callback to 'the sufficiency failure CLAUDE.md already records for the quiet gate versus the llm-pause.' gzip-NCD reads this as high-density overlap with CLAUDE.md's own text even though CLAUDE.md itself was never edited this commit.

    The semantic pull: Literal strings doing the pulling: 'the sufficiency failure CLAUDE.md already records', 'TWO AXES, NEVER MERGED', 'EVERY OVERRIDE CARRIES A TTL', and the header cadence itself (short, bold, all-caps declarative sentence followed by justification) — this is CLAUDE.md's house style being reproduced in a commit message about scripts/ops/, dragging mass into architect's CLAUDE.md-owned coordinate while the 205 real inserted lines sit in an unowned ops-tooling path.

    Intent–reality gap: The commit claims (via Originating-Terminal + Story trailer) that architect 'owns the rung, and the honest note that a categorical envelope was never going to catch it' — but the receipt measures 100% of the actual work landing in scripts/ops/ and tests/ops/, files architect's glob does not claim and that this commit never connects to a CLAUDE.md edit.

    • Add a 'THE SPEND CAP' bullet to CLAUDE.md (STANDING TRIPWIRES / Operational Safety) mirroring THE QUIET GATE's structure — this is architect's actual missing deliverable
    • Add scripts/ops/*.mjs + tests/ops/*.test.mjs to 🎩 operator's owned-surface glob in .workflow/rooms/operator.html (autonomy.mjs, quiet-gate.mjs, wip.mjs, spend-cap.mjs all belong to one ledger family)
    • Add a guard asserting a commit's top Relevant-Room actually owns ≥1 changed file per its .workflow/rooms/*.html glob, OR the commit touches CLAUDE.md when architect is claimed
    • Cross-reference the new CLAUDE.md spend-cap bullet from the existing quiet-gate/llm-pause bullets so future reef ingestion sees one consistent three-axis family instead of one documented axis and two undocumented ones
    • Extract the shared 'ADVISORY — nothing on one machine can stop a process that does not ask' boilerplate (WIP claims, quiet gate, spend cap) into one referenced fragment rather than re-typing it near-verbatim in three tools' headers

    What we learned: A commit message written in CLAUDE.md's own rule-authoring register (bold ALL-CAPS headers, 'Guarded by' citations, direct callbacks to named CLAUDE.md sections) routes to architect via prose-style overlap even when the diff never touches CLAUDE.md and lands entirely in an unowned ops-tooling path — the reef is keying off REGISTER, not path, and a brand-new tool family (scripts/ops/) can be 100% correctly built and still guarantee an out-of-lane receipt because it has no room glob to land in at all.

    The missing guard (the deliverable): tests/ops/room-trailer-matches-surface.test.mjs — fails a commit whose Originating-Terminal/Relevant-Rooms trailer names a room X when zero changed files match room X's .workflow/rooms/*.html owned-surface glob AND CLAUDE.md is not itself part of the diff.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized in volume but truncated in reach on both intent and reality sides (walkTruncated=true, near-equal lit counts) — the walk is hitting its budget ceiling before it can resolve whether the CLAUDE.md-style prose genuinely converges on architect's lattice region or elsewhere. Adjustment: raise --max-depth for commits whose message carries high density of direct-quote citations to another file (a cheap proxy: count of quoted phrases matching existing corpus text), since those are exactly the commits most likely to need extra hops to disambiguate 'citing X' from 'being X'. Experiments: Re-run this commit's reef seed with every literal CLAUDE.md-quoted fragment stripped from the message and compare offPct — isolates how much of the 27% pull is direct phrase reuse versus genuine topical overlap · Seed the reef with the changed file paths (scripts/ops/spend-cap.mjs, tests/ops/spend-cap.test.mjs) as an explicit anchor alongside the message text and measure whether prewalkOverlap rises and offPct falls — tests path-anchoring against prose-anchoring · A/B replay two versions of this same commit message — one keeping the bold ALL-CAPS CLAUDE.md-style headers, one flattened to plain prose describing the identical mechanism — through the offline receipt pipeline to isolate how much placement shift comes from register alone versus content

    Measured outcome (deterministic, not self-graded): room mean offPct 22→42 over 4 commits since · dictionary ablation saved 1133B

    An ops-governance commit that reads like a CLAUDE.md rule but never lands one inflates architect's loss ratio with a false-positive lane claim on every future spend/autonomy commit; premium is restored by shipping the CLAUDE.md entry this prose already earned and re-homing scripts/ops/ under the room that actually operates it.

    Open the signed attestation →

25. 2026-09-05 · 📐 VS Code Architect · offPct 37 · structural

    req(R19): 2,431 single-file commits, 65.5% at sigma<=0, and still not one `unplaced`

    Why it left the lane: docsOnly=true and the diff is a 2-line delta to a JSON tracker (data/requirements/2026-09-05-session-asks.json), so the reality corpus this walk actually ingests is almost entirely the commit MESSAGE prose, not code. That prose is dense PMU-diagnostics vocabulary — 'unplaced sentinel', 'sigma<=0', 'panel path', 'single-file commits', 'sigmaType... drift or spec-delta' — none of which appears in the architect room's declared surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). prewalkOverlap=2 against walksIntent=101 shows the reef seed already found almost no shared vocabulary between the claimed lane and the message text before the walk even ran; the walk then fully saturates both sides (litIntent=101/101, litReality=136/136) and 37% of the 136 lit reality cells land outside the intent lane.

    The semantic pull: The pulling terms are the message's own self-referential PMU vocabulary — 'unplaced sentinel', 'sigma<=0', 'panel path', 'ballistic'-style walk framing, 'vocabulary'/'lattice' self-reference about the pipeline itself. These are laboratory/PMU-internals terms (the Story trailer's own phrase: 'laboratory owns the walk'), not architect-wiring terms, and no line in the diff carries this vocabulary at all — the diff touches only a JSON asks-tracker. So essentially 100% of the semantic mass driving the walk originates in prose describing a different room's territory than the one the commit is filed under.

    Intent–reality gap: The commit message frames this as architect's wiring work ('architect owns the wiring, where this and the last two findings all landed') over a 2-line JSON tracker entry, but the receipt shows 37% of the lit reality mass landed outside that lane — because the R19 sigma/sentinel investigation prose is laboratory's subject matter, not architect's, and it dominates the reality corpus since the diff itself carries no semantic content of its own.

    • Add `data/requirements/*.json` to 🎩 operator's owned-surface glob in `.workflow/rooms/operator.html` so it stops inheriting the originating terminal's lane by default
    • Add a pre-render check in the drift-receipt pipeline: when docsOnly=true, diff LOC is small, and the changed file matches zero room's owned glob, tag the receipt 'UNOWNED-SURFACE' instead of silently attributing lane-fit to the originating terminal
    • Cap walksReality's saturation for docsOnly commits proportional to diff LOC (not message length) so a 2-line JSON delta doesn't inherit a 136-cell walk sized for the prose describing it
    • Run the `recontext` skill over `data/requirements/` specifically to check whether other tracker files (not just this one) share the same unowned-surface gap
    • Update CLAUDE.md's commit-trailer convention section to state the laboratory-vs-architect Relevant-Rooms ordering rule explicitly, since this commit's own Story trailer already drew the distinction the pipeline didn't act on

    What we learned: `data/requirements/*.json` is a cross-room ledger with NO glob owner in any `.workflow/rooms/*.html` file — so a commit touching only that tracker inherits its lane-attribution entirely from the originating terminal, with zero vocabulary check against what the commit is actually about. Because this room's docsOnly commits are typically dense PMU-diagnostic research narratives (laboratory's subject matter), every one of them structurally drifts the same way regardless of who's honest about ownership in the prose — this is a glob-coverage gap, not a wording problem, and it's why 30 events in 14 days haven't converged under prose-only intervention.

    The missing guard (the deliverable): A guard (e.g. `tests/ops/room-surface-coverage.test.mjs`) that fails when a commit's changed-file paths match zero room's owned-surface glob across all `.workflow/rooms/*.html` declarations — surfacing 'unowned path, lane attribution is terminal-only' before the receipt silently prices the commit against whichever room happened to author it.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Over-utilized relative to the work's size: both sides fully saturate (100% lit on both intent and reality) even though the actual diff is a 2-line JSON delta. Adjustment: scale walksReality's budget to diff LOC (or a docsOnly discount factor) rather than to full commit-message length, so a small tracker edit doesn't inherit a 136-cell walk sized for the prose describing it. Experiments: Re-render this sha with intent seeded ONLY from the room's owned-surface glob text (not the full commit message) and check whether prewalkOverlap stays at 2/101 or drops further — isolates glob-vocabulary mismatch from message-vocabulary mismatch · Group the corpus's 2,431 single-file commits by whether their touched file matches ANY room's owned-surface glob, then compare offPct distributions matched-vs-unmatched — tests the ownership-gap hypothesis at corpus scale rather than on one commit · Cap walksReality at walksIntent×1.0 for docsOnly commits with diff LOC<10, re-render this sha, and compare offPct before/after to isolate how much of the 37% is over-saturation versus genuine vocabulary mismatch

    Measured outcome (deterministic, not self-graded): room mean offPct 23→25 over 4 commits since

    A 30-events-in-14-days out-of-lane rate for this room prices its lane at a high loss-ratio the underlying code doesn't deserve, because the receipt currently can't distinguish 'legitimate cross-room research narrative in a tracker commit' from 'actual wiring drift'; premium is restored by giving data/requirements/*.json an explicit owner and de-saturating the docsOnly walk so the receipt starts pricing the two cases differently instead of flagging every research write-up identically.

    Open the signed attestation →

26. 2026-09-05 · 🎤 voice · offPct 25 · structural

    feat(iamfim): show them the blurry graph first, then rerun the query in front of them

    Why it left the lane: The commit's prose ("THE BEFORE-GRAPH IS THE PITCH", "tell the user this is what your graph looks like") reads exactly like voice-room content-craft — persuasive, reader-facing narrative — but every line the diff actually commits is application state and wiring: `const [before, setBefore] = useState(...)`, `if (data?.located) setBefore(...)`, the `fetch('/api/scan-origin', ...)` call inside `onGranted`, and the new lightbox copy block wired to `carrierPlace`/`near`/`total` props in ScanNeighbourhoodAsk.tsx. None of that touches voice's owned surface (`books/tesseract/chapters/*.md`, `src/content/blog/*.mdx`, `decentralizedtalk/slideData.ts`, `src/app/6needs/`). The reality mass landed entirely in `src/components/analysis/*` — a React-state/API coordinate the commit's own Relevant-Rooms trailer already correctly assigns to 🔨 builder and 📐 architect, not 🎤 voice.

    The semantic pull: The pull is narrative vocabulary dragging a code-lane commit toward voice: "the reader is not asked to imagine", "honest in the direction that costs us", "the strongest thing on this page is now an admission" — persuasion-craft language that is voice's normal register — sitting directly above `setBefore(`, `data?.located`, and `/api/scan-origin`. The room label followed the PROSE STYLE of the commit message, not the coordinates of the changed files.

    Intent–reality gap: The message claims a content/UX-narrative achievement ("the before-graph is the pitch", written in voice's persuasive register) but the receipt measured the actual work as React component state management and API sequencing in src/components/analysis — a coordinate voice's owned-surface glob never covers.

    • Add the Relevant-Rooms-first resolution rule to the drift-receipt's trailer parser (wherever Originating-Terminal is currently read as THE lane)
    • Update CLAUDE.md's COMMIT MESSAGE STRUCTURE section with the sentence in spec_rewrite
    • Backfill: re-score the last 108 voice-room off-lane events using Relevant-Rooms instead of Originating-Terminal and report how many collapse to in-lane — this either confirms or kills the mistagged-lane hypothesis before more code changes ship
    • Add a CHANGELOG-visible note to whichever room config drives owned-surface globs (e.g. `.workflow/rooms/voice.html`) stating that physical-terminal commits on feature code are expected and should not count against voice's lane score once the resolver fix lands
    • Add a regression fixture: a commit with Originating-Terminal=voice and Relevant-Rooms=[builder, architect] over files matching builder's glob must resolve to in-lane post-fix (guard-seen-to-fail style: red before the fix, green after)

    What we learned: The room's owned-file-surface glob and the commit-trailer's Relevant-Rooms field are two independently-maintained sources of truth for 'what lane is this commit in', and the drift-receipt pipeline has apparently only ever consulted one of them (Originating-Terminal) — even though the commit-message convention was explicitly designed to let the two diverge. 108 off-lane events in 14 days for one room is the signature of a resolver reading the wrong field, not 108 separate misjudgments by the operator.

    The missing guard (the deliverable): A fixture test asserting: given a commit whose Relevant-Rooms trailer names rooms whose owned-surface glob DOES cover 100% of the changed files, while Originating-Terminal's owned-surface glob covers 0%, the drift-receipt's lane resolution must select Relevant-Rooms and report in-lane — not flag the Originating-Terminal room as off-lane.

    Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. Near-SATURATED: litIntent=141/144 and litReality=129/144 with walkTruncated=false means both sides lit almost the entire 144-cell lattice with no time-budget cutoff. At that density, in-lane and out-of-lane cells are barely distinguishable — the walk has flattened its own discriminative signal. Adjustment: for commits whose reality corpus is dominated by .tsx/.ts identifiers (a feature-code fingerprint distinct from voice's markdown corpus), reduce --max-depth by one ply so the walk halts before covering >90% of the lattice, restoring separation. Experiments: Re-run the last 108 voice-room off-lane events through a Relevant-Rooms-first resolver and measure what fraction collapse to in-lane — confirms or kills the mistagged-lane hypothesis with a number instead of a guess · Correlate prewalkOverlap against final offPct across the last 50 commits per room; if low prewalkOverlap (<10) predicts high offPct with strong correlation, add a pre-walk early-warning flag before spending the full recursive walk budget · A/B the --max-depth ply cap on commits with a .tsx-dominant reality corpus (like this one) vs the current default, and measure whether litReality drops meaningfully below 129/144 while offPct becomes more discriminative (moves away from the saturated near-ceiling reading)

    Measured outcome (deterministic, not self-graded): room mean offPct 24→17 over 16 commits since · dictionary ablation saved 866B

    An off-lane event driven by a resolver bug rather than an actual scope violation inflates this room's measured loss ratio without a real underlying defect, so the fix that restores accurate pricing is making the receipt read Relevant-Rooms before Originating-Terminal — not asking the operator to change terminals.

    Open the signed attestation →

27. 2026-09-05 · 🔨 builder · offPct 36 · mistagged

    fix(mirror): refuse a public push carrying a third party's NAME, not just their address

    Why it left the lane: The intent lane is builder's owned surface (src/app/, src/components/, src/lib/, public/ — UI/component vocabulary). The diff's actual mass sits entirely in scripts/mirror-to-public.sh, scripts/ops/mirror-name-gate.mjs and tests/ops/mirror-name-gate.test.mjs — zero files touch any builder glob. The diff's own vocabulary is security/leak-prevention ("REFUSE A PUBLIC PUSH", "AN ADDRESS IS DECIDABLE, A NAME IS NOT", "watchlist is derived at run time and never committed", "redact()", "ABORT"), which is vault/operator territory, not builder's. Tellingly the commit's own Relevant-Rooms trailer already names 🔒 vault, 📐 architect, ☕ network — builder isn't even on that list — so the drift the receipt measured (prewalkOverlap only 7/140 seeds) is the reef correctly failing to find builder-shaped vocabulary anywhere in this diff.

    The semantic pull: scripts/ops/*.mjs + tests/ops/*.mjs paths and the security-gate lexicon ("third party's NAME", "denylist the path", "private markers", "redact", "ABORT") pulled the reality lattice toward vault/operator coordinates while the Originating-Terminal trailer declared the intent lane as builder — the two never had a chance to overlap.

    Intent–reality gap: The commit message claims this is builder-room work (Originating-Terminal: 🔨 builder), but the receipt measured reality landing 36% outside that lane, and the commit's own Relevant-Rooms trailer independently names vault/architect/network — the work is correct, the room claim on it is wrong.

    • Add `scripts/ops/mirror-*.mjs` and `tests/ops/mirror-*.test.mjs` to vault's (or architect's) owned-surface glob in `.workflow/rooms/*.html` so future commits like this one attribute correctly at ingest time.
    • Extend the commit-trailer helper (`scripts/open-room-session.sh --print-label` or wherever Originating-Terminal is stamped) to warn when the diff's file list matches zero globs on the declared room.
    • Cross-reference this fix with `.workflow` now being denylisted in `scripts/mirror-to-public.sh` — confirm no other dashboard-shaped path (e.g. `docs/decks/**` per-person decks) carries the same bare-name-token risk.
    • File the still-open book-occurrence (Bryan_Lemster/Halcyon worked example) as a tracked human-review item, since the commit explicitly declines to fix it.

    What we learned: Originating-Terminal and file-relevance are allowed to diverge by design (CLAUDE.md says so), but nothing currently flags the degenerate case where the divergence is TOTAL — 0 of N diff files match the declared room's globs while the commit's own Relevant-Rooms trailer names the correct room in the same message. The receipt caught what the trailer convention alone couldn't.

    The missing guard (the deliverable): A post-commit check that computes overlap between the diff's file list and the Originating-Terminal room's owned-surface globs; if overlap is 0 and Relevant-Rooms names a different room, print a WARN (not a block) suggesting the trailer be corrected — cheap, mechanical, and would have caught this exact commit.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated — 128-140 lit cells on each side with no walkTruncated flag. The lever here isn't walk depth, it's reef seed quality: the intent corpus for this commit should ingest the Relevant-Rooms trailer text alongside the message body, since that trailer already names the correct room and would raise prewalkOverlap directly. Experiments: Experiment: re-run this commit's reef ingest with the Relevant-Rooms trailer folded into the intent corpus (not just the commit message body) and measure whether prewalkOverlap rises above 7/140. · Experiment: seed scripts/ops/ and tests/ops/ path tokens into the reef's room-vocabulary bags for vault/architect and re-measure offPct on this commit and the room's other 71 recent out-of-lane events to see how many were the same ops-vs-builder confusion. · Experiment: add the zero-overlap-with-declared-room detector as a dry-run over the last 30 days of commits and count how many would have fired, to size whether this is a one-off or a systemic builder mislabel.

    Measured outcome (deterministic, not self-graded): room mean offPct 20→17 over 4 commits since · dictionary ablation saved 960B

    A security gate this precise (measured false-positive rate, redaction verified by a firing test) lowers the leak-lane's loss ratio regardless of room mistag, but the mistag itself is noise in the room-attribution ledger — fixing the glob ownership restores clean room-level premium signal without touching the gate's own soundness.

    Open the signed attestation →

28. 2026-09-05 · 📐 VS Code Architect · offPct 25 · structural

    ops: one paste for the three things that need a human, in order of leverage

    Why it left the lane: The diff strings that carry the mass are not architecture-strategy vocabulary: 'InMail', '1,706 chars', 'LinkedIn', 'R10, R11 and R13', 'constituent-of-two', 'mirror-to-public.sh --push', '5,253 receipts across 6,111 commit shas', 'a GitHub timestamp — a date we could not choose', 'milestone A4', and 'chrome-cdp.mjs --login'. Every one of those is network/operator subject matter (a send, a publish event, a login credential) wrapped in a bash echo block. The commit's own Story trailer says it outright: 'operator owns all three. network owns the send and the dated event.' The reef correctly grips that vocabulary toward the operator/network coordinates; the intent lane declared by the room (architect / docs/strategy, docs/03-product) never claimed that content.

    The semantic pull: Three concrete pulls: (1) 'InMail'/'LinkedIn'/'constituent-of-two' → ☕ network's send-and-engagement coordinate; (2) 'mirror-to-public.sh --push'/'5,253 receipts'/'GitHub timestamp' → 🎩 operator's irreversible-publish coordinate; (3) 'chrome-cdp.mjs --login' → the ops/session-restore coordinate neither room's glob names. All three sit inside a single new file under docs/ops/, a directory absent from the architect room's owned-surface list entirely.

    Intent–reality gap: The commit message claims a scoped architect artifact ('the handoff, which is the last thing that could be built without him') but its own Relevant-Rooms trailer and body content are majority operator+network subject matter; the receipt measured exactly that split as 25% off-lane mass.

    • Add `docs/ops/**` to 🎩 operator's owned-surface glob in .workflow/rooms/operator.html
    • Update the receipt's lane-computation step to union Relevant-Rooms trailer entries into the intent lane when >1 room is named
    • Add a one-line clarification under DELEGATE-TO-SUBTASKS in CLAUDE.md naming this union-lane rule
    • Run the /recontext skill to confirm docs/ops/ isn't already contested by another room's glob before claiming it for operator
    • Re-render f6e557133's receipt after the glob/lane fix ships and confirm offPct drops toward the 15% threshold

    What we learned: The architect room's 35 out-of-lane events in 14 days are not 35 independent mistakes — this one traces to a single missing structural fact: docs/ops/ is an unclaimed cross-room surface, so any multi-domain handoff script (exactly the artifact DELEGATE-TO-SUBTASKS asks for) is guaranteed to drift because the lane math attributes to the single originating terminal while the commit's own trailer correctly names three rooms. Prior interventions here were facts-only and never named this; this is the first mechanism-level fix available to this room.

    The missing guard (the deliverable): A test asserting: for any commit whose Relevant-Rooms trailer names N>1 rooms, the receipt's intent lane equals the union of those N rooms' owned-surface globs — not the originating-terminal room alone. Paired with a second assertion that every directory touched by 2+ historical commits' Relevant-Rooms trailers is claimed by at least one room's glob (no orphan ops directories).

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent = 1.0 and litReality/walksReality = 1.0 — full saturation, every walked cell lit, combined with walkTruncated=true. This reads as the walk hitting its ply/time-budget ceiling before it could resolve the sparse, multi-domain identifier vocabulary in a shell-script diff (bash + inline node -e). Adjustment: raise --max-depth specifically for diffs where the changed file is a `.sh` wrapper containing an embedded `node -e` block, since that shape packs several coordinate axes into few identifiers and needs more plies to separate them before truncation collapses the spread into a false-positive off-lane reading. Experiments: Ablation: re-seed the reef for this commit using the Relevant-Rooms trailer entries (🎩 operator, ☕ network, 📐 architect) as additional anchor terms, and measure whether prewalkOverlap rises above 3 and offPct drops below 15. · Register `docs/ops/**` as its own ShortLex reef bucket and re-run the last 35 architect-room out-of-lane events to measure how many were docs/ops/ commits vs genuinely mixed-content architect commits. · A/B the lane-union fix: recompute offPct for the last 10 multi-room-trailer commits with union-lane vs single-lane attribution and report the delta in mean offPct.

    Measured outcome (deterministic, not self-graded): room mean offPct 23→32 over 3 commits since · dictionary ablation saved 576B

    An out-of-lane ops-handoff artifact whose own trailer already discloses the multi-room split is a measurement-boundary loss, not a behavioral one — the fix is claiming docs/ops/** and unioning Relevant-Rooms, which restores this lane's premium without touching the underlying (correct) work.

    Open the signed attestation →

29. 2026-09-05 · 🎤 voice · offPct 19

    feat(iamfim): why you actually stopped — name the friction, not the reader

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 21→31 over 6 commits since

    Open the signed attestation →

30. 2026-09-04 · 🎤 voice · offPct 26 · mistagged

    feat(track): /iamfim — visits-per-hour chart + the deterministic human ladder

    Why it left the lane: The entire diff lives under scripts/track/iamfim-visits.mjs, tests/track/iamfim-grader.test.mjs and .claude/commands/iamfim.md — none of these paths match a single one of voice's owned globs (books/tesseract/chapters/*.md, src/content/blog/*.mdx, docs/05-content/blog/scratchpad/, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/, scripts/voice-*-check.sh). The diff's own vocabulary — 'central store', 'user_activity_logs', 'hour-of-day', 'SVG', 'T0/T1/T2/T3 grade ladder', 'HUMAN-80' — is instrumentation/grid vocabulary that has nothing to do with the book/blog prose corpus voice's lane was seeded from. The commit's own Relevant-Rooms trailer already names 🔨 builder, 🎩 operator, 📐 architect and pointedly omits 🎤 voice — the pull toward the out-of-lane coords is confirmed by the commit's own metadata, not just the receipt.

    The semantic pull: scripts/track/*, tests/track/*, user_activity_logs, SVG line-chart rendering, and the T0-T3 grade ladder terms are Operations/Grid-flavored strings with zero lexical overlap against voice's book/blog/6needs corpus — exactly why prewalkOverlap landed at 6.

    Intent–reality gap: The commit claims 🎤 voice as Originating-Terminal, but its own Relevant-Rooms trailer already names builder/operator/architect instead, and the receipt's 26% offPct is just the pipeline independently re-discovering what the trailer already admitted — the work was never voice's to claim.

    • Add 'scripts/track/**', 'tests/track/**' and '.claude/commands/iamfim.md' to 🔨 builder's declared owned-surface glob (rooms.json / the room's owned-surface doc) so future tracking/analytics commits route correctly
    • Re-run sync-vectors.mjs after the glob update and confirm prewalkOverlap for 028a5cea0 rises materially against builder's corpus vs voice's 6
    • Add a pre-commit or post-commit warn that flags when Originating-Terminal is absent from the commit's own Relevant-Rooms list
    • Audit voice's other 97 out-of-lane events in the last 14 days for the same self-contradicting-trailer signature before assuming each needs a bespoke diagnosis

    What we learned: This is the first of voice's four intervention passes where the mistag is self-diagnosable from the commit message alone — Relevant-Rooms and Originating-Terminal already disagreed before the offline reef/walk ever ran. The prior three interventions treated every out-of-lane event as requiring the full pipeline to locate the drift; this one shows a whole subclass (wrong-room attribution on otherwise-correct code) is catchable for free at commit time by comparing two trailers the commit already writes.

    The missing guard (the deliverable): A commit-message linter (pre-commit or a CI check) asserting Originating-Terminal ∈ Relevant-Rooms — refusing or loudly warning when a commit's own two room trailers disagree, which is exactly the signature this event carries.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/litReality near-parity (143/142) alongside walkTruncated=true suggests the walk hit its depth/budget ceiling roughly symmetrically rather than under-walking one side — plausibly right-sized, but the truncation flag means a wider real reality-side spread could still be hidden. Concrete adjustment: re-run 028a5cea0 with --max-depth bumped one tier and compare litReality; if it grows past litIntent, the default depth is truncating the reality side and understating offPct's true source, if it stays ~142 the current depth is adequate. Experiments: Re-run gzip-NCD prewalkOverlap for this sha's diff against all 9 rooms' corpora (not just voice) and confirm builder's overlap is measurably higher than voice's 6 — measurable outcome: builder overlap >20 vs voice's 6 · Add an 'Originating-Terminal ∈ Relevant-Rooms' self-consistency feature to the historical corpus and check its correlation with offPct>15 across past events — measurable outcome: a positive, non-trivial correlation coefficient · Add scripts/track/** and tests/track/** to builder's vector_keywords in rooms.json, re-run sync-vectors, then re-render 028a5cea0's receipt — measurable outcome: offPct should drop toward <15 once attributed to the correct room

    Measured outcome (deterministic, not self-graded): room mean offPct 19→35 over 6 commits since · dictionary ablation saved 667B

    The code itself is sound and cheap to insure — the loss-ratio hit is purely attribution noise inflating voice's already-elevated off-lane rate (97 events/14d), restored by moving the claim (trailer fix + owned-surface glob) rather than the code, so the lane's premium re-prices correctly against builder instead of voice.

    Open the signed attestation →

31. 2026-09-04 · 🎤 voice · offPct 22 · structural

    fix(license): webhook fails closed + idempotent mint — the key email fires once, and only for Stripe

    Why it left the lane: scripts/pmu/drift-watch.sh:27-36 resolves the receipt's ROOM purely from the `Originating-Terminal:` trailer (a case statement matching *Voice* → room=voice) and never reads `Relevant-Rooms:` at all. commit-drift.mjs's own comment says it plainly: '#14: the TO-ROOM drives reef-spec selection... the reef it is graded against is the spec delegated to that room... you don't pass a reef by hand, the room you're in picks it.' So this commit — 57 lines of src/app/api/webhooks/stripe-iamfim/route.ts (webhookSecret, constructEvent, stripe_ref, mintLicense) plus 123 lines of tests/license/*.mjs — got graded against 🎤 voice's reef because the operator's trailer said `Originating-Terminal: 🎤 voice`, even though the SAME commit's own `Relevant-Rooms: 🔒 vault, 🔨 builder, 📐 architect` trailer correctly names the actual content domain. Zero of the three changed files match any of voice's owned globs (books/tesseract/chapters/*.md, src/content/blog/*.mdx, docs/05-content/blog/scratchpad/, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/, scripts/voice-*-check.sh) — the mass had nowhere honest to land inside voice's lane, so 22% of the 99 lit reality cells fell outside it.

    The semantic pull: Terms like 'webhook', 'STRIPE_FIM_IAM_WEBHOOK_SECRET', 'constructEvent', 'mintLicense(', 'stripe_ref', 'handlePerAgentLicense', 'dedupe', 'fail closed' are payments/security/backend vocabulary with no purchase in a content-writing reef; the one line that IS content-shaped (purchase-email.ts's buyer-facing prose about reproducibility) is the sole piece that could legitimately have found voice's lane, and it is a 1-line, 2-word edit against 167 total inserted lines — nowhere near enough mass to pull offPct back under threshold. (No per-region coordinate narrative was cached for this sha, so this is argued from the file-glob/vocabulary evidence available rather than the encircled coords.)

    Intent–reality gap: The commit message and trailers correctly self-declare this as vault/builder/architect work (security fix + idempotency guard on a payments webhook) — the operator explicitly wrote `Relevant-Rooms: 🔒 vault, 🔨 builder, 📐 architect`, deliberately NOT listing voice. But the receipt pipeline ignores that trailer for lane purposes and grades the commit against 🎤 voice's reef solely because that was the terminal the operator physically sat in, producing an offPct spike for content that was never claimed to belong to voice in the first place.

    • scripts/pmu/drift-watch.sh:27-36 — resolve ROOM from `Relevant-Rooms:` trailer first, `Originating-Terminal:` second, 'builder' last; reuse parseRelevantRooms() from delegation-email.mjs instead of a second regex
    • Add src/app/api/webhooks/**, src/lib/license/**, tests/license/** to 🔒 vault's (or 🔨 builder's) owned-surface glob in .workflow/rooms/*.html so future commits in this domain resolve cleanly regardless of which room graded them
    • Add a firing-test guard: a synthetic commit with Originating-Terminal=voice and Relevant-Rooms=[vault,...] must resolve ROOM=vault in drift-watch.sh, both directions (smoke: old code stays voice; no-smoke: a commit with no Relevant-Rooms trailer still falls back to Originating-Terminal)
    • Pull the room's 96 out-of-lane events (14d) and count how many have a changed-file set with zero overlap against voice's owned globs — that count is the blast radius this single fix retires
    • CLAUDE.md § COMMIT MESSAGE STRUCTURE — add the clarifying line named in spec_rewrite so the next session doesn't re-derive this from scratch

    What we learned: The room-resolution step that decides WHICH reef a commit is graded against (commit-drift.mjs #14: 'the room you're in picks it') reads ONLY the Originating-Terminal trailer and structurally ignores the Relevant-Rooms trailer that exists specifically to record content-domain divergence — so every commit made from a terminal that doesn't match its own content (a common, sanctioned pattern per CLAUDE.md's own commit-structure section) manufactures false offPct drift by construction, not by any actual placement error.

    The missing guard (the deliverable): tests/pmu-simulator/room-resolves-from-relevant-rooms.test.mjs — feed drift-watch.sh a synthetic commit message with Originating-Terminal=X and Relevant-Rooms=[Y,...] (Y≠X) and assert the resolved ROOM is Y; feed one with only Originating-Terminal=X and assert ROOM falls back to X. This is the class of guard CODE DEFINITION-OF-DONE demands — the deliverable is this regression test, not the one-line fix alone.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not discriminating: both sides show 100% fill (every walked cell lit, on both intent and reality). That means the walk found no boundary within its own budget — it never had to decide 'this doesn't belong here', it just kept lighting. Concrete adjustment: once the room-resolution fix lands, re-run this sha's walk and check whether fill drops below 100% against the correctly-selected vault/builder reef — if it stays saturated, `--max-depth` may need a decay/null threshold so non-matching cells stop lighting rather than defaulting to lit. Experiments: Re-run sha 067ea85adb's commit-drift.mjs with --room vault (bypassing drift-watch.sh's resolver) and compare offPct/litReality against the shipped voice-room run — if offPct drops materially, that is the direct measured confirmation of this diagnosis · Ablate prewalkOverlap: rerun the reef seed step restricted to each of voice's owned globs only (vs whole-repo) across this room's last 20 commits, and correlate prewalkOverlap with offPct — test the hypothesis that low prewalkOverlap predicts high offPct specifically for zero-glob-overlap commits · After the drift-watch.sh fix ships, track room-mean offPct for 🎤 voice over the next 10 commits and compare against the last 3 events' measured 27→39, 0→36 trend — a fix that actually addresses the mechanism should show voice's mean offPct fall toward its true content-drift baseline rather than continuing to spike on unrelated backend work

    Measured outcome (deterministic, not self-graded): room mean offPct 20→34 over 8 commits since · dictionary ablation saved 743B

    A room whose measured offPct is inflated by a mislabeled-lane artifact rather than real content drift is mispriced risk — this lane's premium should fall (or the room excluded from this event's loss-ratio entirely) once the room-resolution fix and its firing-test guard ship and the next batch of voice-room commits shows the false-positive signal gone.

    Open the signed attestation →

32. 2026-09-04 · 📐 architect · offPct 56 · structural

    chore(tree): land in-flight drafts, deck bodies and book-chapter edits

    Why it left the lane: The diff's dominant textual mass is not docs/strategy or docs/03-product content (architect's owned surface) — it's three book-chapter metavector blocks (chapter-03, chapter-04, chapter-08) repeating '**Meta vector — what this section's idea rests on, and what rests on it.**' plus new chapter-08 prose about 'Oja's rule' and 'Hebbian Wiring', and the rewritten afternoon_organizer.html carrying named outreach content ('Patrick Tiernan, Lloyd's', 'Juergen Reinhart, Munich Re', 'Vanessa Maxwell, Allianz', 'Katya Andresen, Cigna', 'Giovanni, CRS', 'Lux Capital'). Every one of those terms pulls the reef toward book/voice/comms coordinates outside architect's five declared globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md). No region narrative is cached for this sha, so exact [row,col] anchors aren't available, but the file-level evidence alone accounts for the 56% off-lane share: of the 144 files, the semantically heavy ones (book chapters, organizer html, blog scratchpad txt) all sit outside the room's owned surface.

    The semantic pull: 'Meta vector' / glossary-address boilerplate (repeated verbatim 3x), 'Verification Cost, Priced Per Crossing (k_E × Volume × Wage)', 'Oja's rule', 'Hebbian Wiring' pull toward book/pedagogy coordinates; 'Lloyd's', 'Munich Re', 'Allianz', 'Cigna', 'CRS', 'Lux Capital', 'aiSure', 'the Omnibus delayed high-risk' pull toward outreach/comms coordinates. Neither cluster shares tokens with architect's own lane vocabulary (strategy/product docs), so the reef mass is pulled toward whichever rooms' reefs those tokens actually seed — book and voice, not architect.

    Intent–reality gap: The message accurately describes a 144-file rescue-sweep across drafts, deck bodies, cook-round transcripts and three peer-claimed book chapters, and its own Relevant-Rooms trailer already names voice and operator alongside architect — the commit is honest about being multi-domain. The receipt still measured 56% of lit reality mass outside the lane because the scorer only checks Originating-Terminal's owned-surface globs, so a commit that correctly declares itself cross-room is punished exactly as hard as one that lied about its scope.

    • Add the scope-exemption clause to CLAUDE.md's COMMIT MESSAGE STRUCTURE section (spec_rewrite above).
    • Finally ship the unified room-drift-check.sh with union(Relevant-Rooms) scoring for chore(tree)/sweep commits — three rooms already wrote the same spec and never merged it.
    • Move afternoon_organizer.html out of repo root (FILE MANAGEMENT rule: 'NEVER save to root folder') into docs/ops/ or docs/comms/self/.
    • Re-run this sha's walk with a raised --max-depth/time budget and record whether offPct drops once walkTruncated=false (tests the walk-truncation hypothesis in pipeline_probe).
    • Cross-reference the 177 architect out-of-lane events in the last 14 days against how many carry a chore(tree) or large multi-file-sweep shape — if it's a large share, this single fix retires most of that backlog at once.

    What we learned: A commit's Originating-Terminal and Relevant-Rooms trailers can both be filled out correctly and honestly (this commit's Relevant-Rooms already names voice+operator) and the receipt will still fire a hard drift alarm, because the scorer has never been taught to read Relevant-Rooms OR the conventional-commit scope (`chore(tree):` is itself a machine-parseable admission of multi-domain scope) — the fix ff09774e5 already proposed (read Relevant-Rooms) is necessary but not sufficient; the scope keyword is a second, even cheaper signal sitting unused in the same commit subject line.

    The missing guard (the deliverable): A test asserting: given a commit whose subject scope is `tree` (or equivalent sweep marker) and a Relevant-Rooms trailer listing N rooms, the room-lane detector's offPct is computed against union(ownedGlobs(N rooms)) and NOT against Originating-Terminal's globs alone — fired against this exact sha's diff shape (book chapters + comms drafts + blog scratchpad + telemetry state, 144 files) as the real fixture, per the GUARD-SEEN-TO-FAIL discipline.

    Detector probe: signal attribution — gzip reef 20% · ballistic walk 80%. Under-budgeted for this commit's size: litReality (127) already exceeds litIntent (94) and walkTruncated=true means the reality walk hadn't finished converging when it was cut off. For a 144-file commit, raise --max-depth (or the time budget) once changed-file count exceeds a threshold (e.g. >50), so the reality walk reaches its natural stopping point instead of being truncated mid-expansion, which inflates apparent off-lane mass. Experiments: Re-run 07fd254d8's walk with --max-depth doubled and compare offPct against this event's 56% — if it drops substantially, confirms walkTruncated (not room-tagging) is the dominant driver and reprioritizes the fix toward budget, not lane logic. · Seed the reef with union(Relevant-Rooms) coordinates instead of only Originating-Terminal's for chore(tree)-scoped commits, then measure whether prewalkOverlap and the litReality/litIntent asymmetry both shrink. · Tally walkTruncated=true rate across all 177 of this room's 14-day out-of-lane events — if truncation correlates with high offPct broadly (not just here), the budget fix generalizes past this one commit class.

    Measured outcome (deterministic, not self-graded): room mean offPct 23→6 over 7 commits since · dictionary ablation saved 1128B

    A structural false-positive at this scale (56% off-lane on a commit whose trailers were fully honest) pushes the architect lane's measured loss-ratio up on noise rather than real drift, which is the cheapest kind of premium bleed to fix — restoring it is one scoring-logic patch (union-of-Relevant-Rooms for declared sweeps) plus its firing-tested guard, not a change in how the room actually works.

    Open the signed attestation →

33. 2026-09-04 · 🎤 voice · offPct 16 · mistagged

    feat(iamfim): last 24 hours on the wall clock, and each newsletter's own first 48

    Why it left the lane: No region narrative is cached for 0b070b3c2 so the encircled coords themselves aren't nameable, but the vocabulary that dominates the diff is diagnostic on its own: `selectSends`, `ndjson`, `.thetacog/email-sent.ndjson`, `minRecipients`, `subject`, `ts`, `to.length`, row-counting, ledger-parsing, SVG bucket math (`n`, `tick`, `xFmt`). None of that is prose/manuscript vocabulary — it's tracking-pipeline/ops vocabulary. The room's own owned-surface list (books/*.md, blog/*.mdx, decentralizedtalk, 6needs) has zero overlap with the two changed files (scripts/track/iamfim-visits.mjs, tests/track/*.test.mjs), which is exactly the shape offPct=16/threshold=15 measures: reality mass landing outside the intent lane's declared surface. prewalkOverlap=4 (very low, out of 134 intent walks) corroborates it at the reef-seed level — almost nothing in the pre-walk seed matched between voice's declared lane and this commit's actual content.

    The semantic pull: The commit's own trailers pull in two directions simultaneously and the receipt evidently followed the wrong one: `Originating-Terminal: 🎤 voice` (which physical terminal the operator happened to be sitting in) versus `Relevant-Rooms: 🧭 navigator, 🎩 operator, 📐 architect` (which the commit message itself names as the rooms whose owned-surface this actually touches). CLAUDE.md explicitly sanctions that divergence as normal — but if the lane-drift pipeline attributes the commit's room for offPct purposes to Originating-Terminal rather than Relevant-Rooms, a purely ops/tracking commit will always read as 'voice, out of lane' no matter how well-formed the work is.

    Intent–reality gap: The commit message accurately describes what it did — extract selectSends(), fix the to.length/row-count defect, add two panels — and the diff matches that description exactly; there is no gap between message and diff. The gap is between the RECEIPT'S room attribution (🎤 voice) and the work's actual home (navigator/operator/architect, by the commit's own Relevant-Rooms trailer and by the fact that zero touched files sit under voice's owned globs).

    • Locate the commit→room attribution step feeding the per-room offPct trend and add the Relevant-Rooms-aware resolution described in code_rewrite
    • Add scripts/track/** and tests/track/** to whichever of navigator/operator/architect's owned-surface glob currently has no claim on them (confirm via .workflow/rooms/*.html — this commit's Relevant-Rooms names all three, so the gap may be that none of them formally owns it either)
    • Add the CLAUDE.md sentence from spec_rewrite fixing which trailer the receipt pipeline treats as authoritative
    • Add the guard from missing_guard
    • Audit the room's other 101 out-of-lane events from the last 14 days for the same Originating-Terminal≠Relevant-Rooms signature before assuming every one of them is a real voice-lane violation — the prior three interventions' rising offPct trend (27→39, 0→36, 26→20) is consistent with this same mis-attribution recurring uncorrected

    What we learned: The receipt pipeline appears to key a commit's lane-drift measurement off Originating-Terminal (which chair the operator was in) rather than Relevant-Rooms (which surface the work actually touched) — even though CLAUDE.md's own commit-message spec explicitly names these as two different, legitimately-divergent facts. This is a new mechanism relative to the three prior facts-only interventions for this room, none of which diagnosed why voice's offPct kept climbing (27→39, 0→36, 26→20) despite no sensemaking ever pointing at a cause.

    The missing guard (the deliverable): A test that constructs a commit whose Relevant-Rooms trailer names rooms different from its Originating-Terminal trailer, feeds it through the room-attribution step that produces the per-room offPct label, and asserts the resulting lane room is chosen by owned-surface glob match against Relevant-Rooms — never by Originating-Terminal alone.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Both sides hit 100% lit-to-walked ratio (134/134, 131/131) while walkTruncated=true — the walk filled everything it reached but was cut off by the depth/time budget before it could reach further, so it's saturated-and-truncated rather than under-used. Adjustment: re-run this commit's walk with --max-depth raised by 1-2 plies specifically for commits whose diff touches non-owned-surface paths, to see whether the truncation itself is inflating offPct by stopping reality's walk before it could re-converge toward any legitimately shared coordinate. Experiments: Re-run 0b070b3c2's receipt with --max-depth +2 and compare offPct — if it drops materially, walkTruncated was a real contributor and default depth should rise for large/multi-concern diffs · Recompute prewalkOverlap seeding the reef from the Relevant-Rooms trailer's owned-surface globs (navigator/operator/architect) instead of Originating-Terminal's (voice) and compare against this event's offPct=16 — directly tests the mistagged-attribution hypothesis using the pipeline's own numbers · A/B two reef corpora, with and without ops/tracking vocabulary (ledger, row, subject, recipients) weighted into Operations-axis coordinates, and measure whether prewalkOverlap rises for commits like this one

    Measured outcome (deterministic, not self-graded): room mean offPct 23→35 over 6 commits since · dictionary ablation saved 824B

    An offPct event caused by mis-attributed room labeling rather than real scope-creep inflates this lane's loss-ratio on a false signal, which is worse than a true miss because it erodes trust in the whole per-room premium; restoring it means fixing the attribution step so voice's premium reflects only commits that actually touched its owned surface.

    Open the signed attestation →

34. 2026-09-04 · 🎤 voice · offPct 40 · mistagged

    feat(track): the clock-contradiction panel — browser clock vs IP clock, Δ-histogram + mismatch ledger

    Why it left the lane: The diff's entire vocabulary — `tzOffsetMinutes`, `clockMismatchHours`, `Intl.DateTimeFormat`, `clockDelta`, `deltaHist`, `mismatchRows`, living in `scripts/track/iamfim-visits.mjs` and `tests/track/iamfim-grader.test.mjs` — is tracking-pipeline engineering vocabulary with zero lexical overlap to voice's owned surface (`books/tesseract/chapters/*.md`, `src/content/blog/*.mdx`, `decentralizedtalk/slideData.ts`, `src/app/6needs/`). Tellingly, the commit's own `Relevant-Rooms` trailer already names 🔨 builder, 🧪 laboratory, 🎩 operator — and pointedly omits 🎤 voice. The intent lane the receipt built from the message pulled toward builder/laboratory-shaped code coordinates; only the `Originating-Terminal: 🎤 voice` line ties this event to voice's ledger at all.

    The semantic pull: Terms like 'panel', 'histogram', 'ledger', 'guard', 'test' are generic scaffolding words that appear in every room's commits, so they don't anchor to voice specifically — but the load-bearing nouns ('client_tz', 'IP timezone', 'VPN', 'beacon', 'route.ts') are pure analytics-pipeline nouns that pull straight at builder/laboratory's owned surface, dragging offPct to 40 against voice's book/blog lane.

    Intent–reality gap: The receipt scores this event as room 🎤 voice's work (via Originating-Terminal), but the commit message's own Relevant-Rooms trailer already states the true owners are builder/laboratory/operator — the commit is honest about its own domain, the lane-scorer just isn't reading that field.

    • Update the room lane-fit scorer (wherever offPct is computed per Originating-Terminal) to fall back to Relevant-Rooms' owned-surface union when Relevant-Rooms excludes Originating-Terminal
    • Add `scripts/track/*.mjs` and `tests/track/*.test.mjs` to builder's (or laboratory's) `.workflow/rooms/*.html` owned-surface glob — the commit's own trailer already claims this territory
    • Backfill: re-score voice's last 14 days of out-of-lane events for the same Relevant-Rooms≠Originating-Terminal mismatch before trusting the room-mean-offPct trend (98 events is suspiciously high for a content room)
    • Note in `scripts/room.sh` or the dictation-loop docs that a voice-terminal session producing non-content work should either bifurcate to the correct room before committing, or the commit should be scored against its declared Relevant-Rooms, not its terminal of origin

    What we learned: Originating-Terminal encodes input modality (which app the operator was dictating into), not output domain (which room's owned surface the diff touches) — for 🎤 voice specifically this is a chronic, structural mismatch because dictation-driven sessions route arbitrary cross-domain work through whichever terminal was open, and the commit's own Relevant-Rooms trailer already carries the correction that the lane-scorer never reads. This is new: the three prior lessons for this room shipped with no sensemaking at all (facts-only), so this is the first mechanism named for voice's 98-events/14-days out-of-lane rate.

    The missing guard (the deliverable): A guard asserting: for any commit where Relevant-Rooms excludes Originating-Terminal, the receipt's lane-fit (offPct) must be computed against Relevant-Rooms' owned-surface union, with a firing test seeded from this exact commit (0d3316bd0) — clean tree scores voice's raw owned-surface (should fail lane-fit), corrected tree scores against builder/laboratory's surface (should pass).

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not shallow: litIntent/walksIntent and litReality/walksReality are both ~100% (137/137, 138/138), and walkTruncated=true confirms it hit its depth/budget ceiling while still fully lit. Adjustment: for commits with prewalkOverlap below a low threshold (e.g. <5), cap max-depth tighter rather than letting the walk run to full budget — a low-overlap seed saturating at full budget just amplifies the seed mismatch across more cells instead of correcting it. Experiments: Re-run the reef seed for this commit scoped to builder's owned-surface corpus instead of voice's, and measure whether prewalkOverlap rises materially (tests whether the low overlap is a room-mismatch artifact vs a genuine content-sparsity issue) · Ablation: strip generic scaffolding strings ('Guard:', 'NO SMOKE —', 'SMOKE —') from the intent corpus before walking on a batch of 20 recent commits, and measure whether litIntent count drops and prewalkOverlap becomes a better mistag predictor · Compute prewalkOverlap across all commits where Relevant-Rooms≠Originating-Terminal vs where they match, and test whether prewalkOverlap<5 correlates with offPct>threshold strongly enough to serve as a cheap pre-walk mistag flag before the full walk runs

    Measured outcome (deterministic, not self-graded): room mean offPct 23→17 over 7 commits since · dictionary ablation saved 923B

    Scoring this commit's tracking-pipeline work against voice's book/blog lane inflates voice's apparent loss-ratio with drift that isn't its own, mispricing that lane's premium; restoring accuracy means scoring lane-fit against the commit's self-declared Relevant-Rooms rather than its Originating-Terminal.

    Open the signed attestation →

35. 2026-09-04 · 🔨 builder · offPct 23 · structural

    fix(initiative): authorship outranks recency, and a proposal's ceiling is draft

    Why it left the lane: The diff's own vocabulary is governance-of-autonomy language, not construction language: `isProposal`, `authorRank`, `denyClass`, `verdict = 'draft'`, and the comment 'the pipeline authorizing itself from its own output, which is the closed loop record conservation says cannot certify itself' are all ratchet/ceiling terms — the exact register of the AUTONOMY LEVEL IS THE RATCHET rule, which is operator-domain vocabulary. `initiative-mine.mjs`'s new `authorRank()` also sorts across 'gemini-web', 'iMessage span', 'spine' sources — cross-channel routing, which is navigator territory. None of that vocabulary has any overlap with 🔨 builder's owned surface (src/app, src/components, build-book-html.sh, decentralizedtalk/, thetacoach-crm-mcp/), so the reef seed and the walk both pulled mass toward the operator/navigator coordinates while the receipt scored the commit against builder's lane.

    The semantic pull: scripts/ops/initiative-frame.mjs and scripts/ops/initiative-mine.mjs carry zero terms from builder's construction/UI vocabulary and carry the operator ratchet's own words back at it ('ceiling', 'draft', 'act', 'authorizing itself from its own output') — the same words CLAUDE.md itself uses for the AUTONOMY LEVEL rule.

    Intent–reality gap: The commit trailer claims Originating-Terminal 🔨 builder, and the receipt scored it against builder's lane, but the commit's own Relevant-Rooms line already names 🎩 operator and 🧭 navigator as the true domains — the message is more honest than the lane it was measured against.

    • Register scripts/ops/** (or at minimum initiative-*.mjs) under 🎩 operator's owned-surface glob in .workflow/rooms/operator.html
    • Re-run the lane/receipt check for 0df45d0889 after the glob update and confirm offPct falls under 15 — this is the falsifier for the structural hypothesis, not just an assertion
    • Grep .workflow/rooms/*.html for 'scripts/ops' across all 9 rooms to find how many other files in that directory are currently unowned by construction
    • Add a one-line note near the COMMIT MESSAGE STRUCTURE section of CLAUDE.md stating which trailer field (Originating-Terminal vs Relevant-Rooms) the lane-threshold detector actually scores against, since the current text permits divergence without saying which side is authoritative for offPct
    • Run scripts/room.sh builder and confirm scripts/ops/initiative-*.mjs does NOT surface in its briefing — if it does, the room-detection layer itself needs the same fix as the glob registry

    What we learned: scripts/ops/ has grown into a real, expanding pipeline-governance surface (this commit's own initiative-mine/frame work, the ratchet, the autonomy ledger) that sits outside all 9 rooms' owned-file globs — every commit that builds there, regardless of which terminal authored it, registers as out-of-lane until the glob registry catches up to where the codebase actually grew. The prior three lessons for this room were facts-only and offPct trended flat-to-worse afterward (16→46, 15→16); this is the first sensemaking pass to find that the room's lane definition itself, not the room's discipline, may be the thing that's wrong.

    The missing guard (the deliverable): tests/ops/owned-surface-coverage.test.mjs — walks every path touched by the last N commits (or a directory allowlist starting with scripts/ops/) and fails if it matches zero room globs across .workflow/rooms/*.html, so a newly-born pipeline directory can't silently accumulate unowned commits the way scripts/ops/ apparently has.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent = 128/128 = 100% and litReality/walksReality = 138/138 = 100% — full saturation on both sides, every cell the walk visits gets counted as lit with no discrimination between a strong hit and a marginal one. That means offPct here is really just 'which lane won the ballistic race,' not a graded signal. Adjustment: introduce a lit-threshold (e.g. only count cells above median edge-weight as lit) so the ratio can fall below 100% and the panel regains texture instead of flooding. Experiments: Register scripts/ops/** under operator's owned-surface glob and re-run this commit's receipt — if offPct drops below 15, the unowned-surface hypothesis is confirmed rather than assumed · Re-seed the reef with the scripts/ops/initiative-*.mjs corpus already ingested (rather than post-hoc) and re-measure prewalkOverlap — tests whether the 6% overlap is an ingest-recency artifact or a true semantic mismatch with builder's lane · Re-run the same commit's ballistic walk with --max-depth reduced by one ply and compare litIntent/walksIntent — if the ratio drops below 100%, the walk is oversaturating at the current depth and the threshold-lit fix above is warranted

    Measured outcome (deterministic, not self-graded): room mean offPct 17→20 over 6 commits since · dictionary ablation saved 1013B

    An out-of-lane event on a real, correct, well-guarded governance commit reads as a false positive against builder's loss ratio for a defect that is actually a registry gap; premium is restored the moment scripts/ops/** is registered to its true owner and the coverage guard ships, not by asking builder's commits to behave differently.

    Open the signed attestation →

36. 2026-09-04 · 🔨 builder · offPct 22

    feat(bookclub): both foot banners return — the domain sticker and the hiring sticker

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 20→38 over 4 commits since

    Open the signed attestation →

37. 2026-09-04 · 🎤 voice · offPct 36 · structural

    strategy(positioning): the record set straight — eight canonical rows, our sentence leads, their frame never quoted

    Why it left the lane: The only file in the diff is docs/strategy/governance-layer-valuation-positioning-2026-09-03.md, which is not on 🎤 voice's owned surface (books/tesseract/chapters, src/content/blog, decentralizedtalk, 6needs, voice-check scripts). The prose itself carries insurance/ops/pricing vocabulary — 'the receipt layer is the balance sheet', 'insurability', 'underwriters', 'CG 40 47', 'detected, placed, priced, dispatched', 'exclusion forms' — none of which are voice's blog/book craft nouns, and everything about it points at the coordinate space the commit's own Relevant-Rooms trailer names: 📐 architect, 🎩 operator, 🎭 performer. voice never appears in Relevant-Rooms at all.

    The semantic pull: prewalkOverlap=7 against walksIntent=117 (~6%) shows the reef seed found almost no shared vocabulary between voice's owned-surface corpus and this strategy doc before the walk even started — the divergence was baked in at seed time, not discovered mid-walk. litIntent==walksIntent (117/117) and litReality==walksReality (137/137) both fully saturate, meaning the walk adds no further discrimination once seeded; the 20-cell gap (137 vs 117) that produces offPct=36 is inherited straight from the seed mismatch. No region narrative was cached for this sha, so I'm reasoning from the owned-surface mismatch and the diff's own vocabulary rather than named coords.

    Intent–reality gap: The commit's own Relevant-Rooms trailer names architect/operator/performer as the rooms whose owned surface this touches, and CLAUDE.md's COMMIT MESSAGE STRUCTURE section says explicitly that Originating-Terminal and file-relevance 'can diverge — that's fine and worth recording.' The receipt pipeline nonetheless scored this commit against the voice lane (Originating-Terminal) and flagged 36% of its mass out-of-lane — penalizing exactly the divergence the repo's own convention declares intentional.

    • Add `docs/strategy/**/*.md` to architect's (or operator's) owned-surface glob in `.workflow/rooms/architect.html` — it currently claims no docs/strategy files
    • Patch CLAUDE.md's COMMIT MESSAGE STRUCTURE section per spec_rewrite above
    • Patch the receipt-to-room binding step per code_rewrite above
    • Grep git log for `Originating-Terminal: 🎤 voice` commits touching `docs/strategy/` over the last 14 days to size how much of voice's 99 out-of-lane events are this same lane-assignment artifact vs real content drift
    • Add the missing_guard test below
    • If the backfill in item 4 shows this is systemic across multiple rooms, promote it to a new anti-rule in `docs/architecture/anti-rules-ledger.md` ("originating-terminal is not a lane")

    What we learned: The repo already models the exact divergence the drift pipeline is penalizing — CLAUDE.md's own trailer convention distinguishes Originating-Terminal (where you sat) from Relevant-Rooms (what you actually touched) and calls the gap between them 'fine' — but the receipt's lane computation reads only Originating-Terminal, so every commit typed from one room's terminal about another room's territory scores as drift by construction. Given voice logged 99 out-of-lane events in 14 days, a real fraction of that count is likely this exact artifact rather than genuine semantic drift, and it will keep firing until the pipeline reads the field the repo already writes for this purpose.

    The missing guard (the deliverable): tests/ops/lane-uses-relevant-rooms.test.mjs — construct a synthetic commit whose changed files match zero globs in the Originating-Terminal room's owned surface but match another room's owned surface (e.g. a docs/strategy/*.md file under a voice-originated commit), run it through the receipt-to-room binding step, and assert the row is NOT scored against the Originating-Terminal room's offPct lane (either it scores against the glob-matched room, or it is tagged 'unowned-surface' and excluded from both rooms' running means).

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Right-sized in mechanics (full saturation, walkTruncated=false — it isn't starving for depth) but mis-seeded in scope: litIntent=117 reflects only voice's owned-surface anchors, so reality's 137-cell footprint over strategy content trivially exceeds it. Concrete adjustment: when Relevant-Rooms names additional rooms in the trailer, seed the intent walk from the UNION of those rooms' owned-surface anchors (not just Originating-Terminal's), which should legitimately widen litIntent and let genuine cross-lane collaboration converge instead of reading as drift. Experiments: Re-run this commit's receipt with the intent walk seeded from the union of Relevant-Rooms (architect+operator+performer) instead of Originating-Terminal (voice) alone, and measure whether offPct drops below the 15 threshold — a direct test of the mistag hypothesis · Backfill the last 20 commits with `Originating-Terminal: 🎤 voice` that touch `docs/strategy/**` and compare their offPct distribution against the last 20 voice commits touching `books/tesseract/**` or `src/content/blog/**` — confirms whether unowned-surface files are a systematic offender for this room · Once docs/strategy/**/*.md is added to architect's owned-surface glob, re-render this commit's receipt and confirm prewalkOverlap rises meaningfully above 7, evidencing the seed (not the walk) was the lever

    Measured outcome (deterministic, not self-graded): room mean offPct 24→16 over 10 commits since · dictionary ablation saved 445B

    A lane-scoring defect that penalizes the repo's own documented divergence inflates voice's loss ratio with false positives, which only clears once the pipeline scores against owned-surface glob match (Relevant-Rooms) instead of Originating-Terminal and the backfill confirms how much of the 99-event count was this artifact.

    Open the signed attestation →

38. 2026-09-04 · 🎤 voice · offPct 32 · structural

    fix(voice): the behavior-guarantee guard was flagging its own canonical sentence

    Why it left the lane: 36 of 37 changed lines sit inside tests/voice/no-behavior-guarantee.test.mjs and are pure test-infrastructure vocabulary — `REFUTES`, `\b(...)\b`, `new RegExp(BEHAVIOR_GUARANTEE.source, 'i')`, `refutedNearby(`, `assert.ok(` — not the prose/content vocabulary (blog, book, dispatch copy) the room's declared lane is built from. No region narrative is cached for this sha, so I can't cite specific [row,col] coords, but the aggregate numbers corroborate the pull: prewalkOverlap=5 against litIntent=119 (~4% seed overlap) means the reef already planted intent (commit-message prose) and reality (regex/test code) in nearly disjoint neighborhoods before the walk even started, and reality's walk covers 21 more cells than intent's (140 vs 119) — the code-vocabulary side simply has more surface to spread over once seeded apart.

    Intent–reality gap: The commit message frames this as a `fix(voice)` correction to a voice guard's regex — content the room legitimately owns by subject matter. The receipt measures 32% of the landed mass outside the intent lane because the entire diff lives in tests/voice/no-behavior-guarantee.test.mjs, a path the room's own owned-surface glob (books/*.md, blog/*.mdx, decentralizedtalk/, scripts/voice-*-check.sh) never claims — the work is voice work by subject, but not voice-owned by path.

    • Add `tests/voice/**/*.test.mjs` to the 🎤 voice room's owned-surface glob in `.workflow/rooms/voice.html`.
    • Audit the other 8 rooms' owned-surface globs for the same shell-check/test-check split (any room with `scripts/<room>-*-check.sh` claimed should also claim `tests/<room>/**/*.test.mjs`).
    • Re-render this commit's receipt after the glob update and confirm offPct drops toward the 15 threshold — the measured before/after is the actual evidence, not this hypothesis.
    • Add a parity guard (see missing_guard) so a future under-scoped room glob fails loudly instead of silently generating out-of-lane events for 14 days at a time.
    • Note in the guard's own header comment that its home directory (tests/voice/) is a claimed room surface, so the next person editing it doesn't rediscover this gap.

    What we learned: The room's owned-surface declaration only lists the shell wrapper (`scripts/voice-*-check.sh`) for its own guards, not the `tests/voice/**` directory those wrappers actually call into — so every guard-only fix in this room's test suite is structurally guaranteed to read as out-of-lane, regardless of how correct or well-scoped the fix is. This is a new failure mode versus the prior three lessons for this room (which were facts-only, no diagnosed mechanism) — it's not a one-off tagging slip, it's a standing gap in the glob that would recur on every future tests/voice/*.test.mjs commit until the glob itself is patched.

    The missing guard (the deliverable): A room-surface parity test (e.g. `tests/ops/room-surface-parity.test.mjs`) that, for every room declaring a `scripts/<room>-*-check.sh` wrapper in its owned-surface glob, asserts the room also claims `tests/<room>/**/*.test.mjs` — failing loudly at glob-declaration time instead of leaking out as dozens of downstream out-of-lane receipts.

    Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. litIntent/walksIntent and litReality/walksReality are both 100% (every walked cell lit), which reads as saturated rather than shallow — but walkTruncated=true means the walk hit its depth/time budget before natural termination while still lighting every step, so the measured 140-cell reality spread is a floor, not the walk's true extent. Concrete adjustment: bump `--max-depth` (or the time budget) specifically for commits whose diff is >90% inside a `tests/**` path, since regex/test vocabulary tends to have denser reef connectivity and saturates the default budget faster than prose commits do. Experiments: Re-render sha 113b434ce5's receipt after adding `tests/voice/**/*.test.mjs` to the room glob and measure whether offPct falls under 15 — direct test of the spec_rewrite hypothesis. · Re-run the same commit's walk with `--max-depth` doubled and compare litReality/walksReality to the current 140/140 — if the walk is truly saturated pre-truncation, extending the budget should reveal additional cells and a higher true spread, testing whether walkTruncated=true is masking the real offPct. · Compute prewalkOverlap across the room's last 10 commits, split by whether the diff touches `tests/voice/**` vs `src/content/blog/**` or `books/**`, to confirm the ~4% seed-overlap pattern seen here is systematic to test-file commits and not an artifact of this one sha.

    Measured outcome (deterministic, not self-graded): room mean offPct 22→18 over 15 commits since · dictionary ablation saved 1230B

    A structurally under-scoped owned-surface glob generating an out-of-lane receipt on a genuinely correct, well-tested guard fix is a false loss signal — it inflates this lane's apparent claim frequency without any real deviation, so the fix (widen the glob, add the parity guard) restores the premium basis by removing noise rather than by changing behavior.

    Open the signed attestation →

39. 2026-09-04 · 🎤 voice · offPct 25 · mistagged

    fix(track): persist client_tz + utm from the visit beacon — the fields it always sent, the insert always dropped

    Why it left the lane: The diff's semantic mass is dominated by data-plumbing nouns — 'client_tz', 'utm', 'event_data insert', 'body.timezone', 'body.utm', 'Object.keys' — none of which are Tactics.Signal/Operations.Grid vocabulary (git-discipline, commit iteration, hook staging). The reef's B3.Tactics.Signal ×11 and C1.Operations.Grid ×5 pulls are the walk finding 'signal'-shaped words (utm IS literally a marketing signal token, client_tz IS a routing/attribution signal) and 'grid'-shaped words (event_data as a row/column insert into a table) and mapping them onto the Tactics/Operations lattice instead of the 🎤 voice room's declared lane (blog/book/content). The commit is track/route/analytics code, not narrative content — its own terms have no home in voice's owned surface, so the walk correctly finds no in-lane anchor and floods the adjacent Tactics×Operations space that DOES recognize 'signal' and 'grid' vocabulary.

    The semantic pull: 'utm' → B3.Tactics.Signal (campaign/attribution signal, direct lexical hit); 'event_data:' insert block + 'client_tz'/'Object.keys' → C1.Operations.Grid (row/column data-grid insert); 'route.ts', 'POST(req...)' → C2/C3 Operations.Loop/Flow (request-handling plumbing); none of these terms appear anywhere in 🎤 voice's owned globs (books/tesseract/*.md, src/content/blog/*.mdx, decentralizedtalk/slideData.ts) so the reef has nothing in-lane to catch the mass — it falls through to the nearest recognizing coordinates.

    Intent–reality gap: The commit message and trailers claim room 🎤 voice (Originating-Terminal), but the entire diff is backend tracking/analytics code (src/app/api/track/visit-shape/route.ts) plus its test — content this room does not own and the receipt correctly measures as landing at C.Operations × B3.Tactics.Signal, not in voice's blog/book lane.

    • Add src/app/api/track/**/*.ts to a room's owned-surface glob (🔨 builder or 🧪 laboratory per .workflow/rooms/*.html) so future commits here score in-lane against the correct room instead of defaulting to Originating-Terminal.
    • Update the lane-fit scorer (wherever offPct/room attribution is computed) to weight Relevant-Rooms file-glob ownership over the Originating-Terminal trailer when the two diverge.
    • Add a CLAUDE.md note under the room-naming section clarifying Originating-Terminal = physical terminal (may legitimately differ from file-lane) vs Relevant-Rooms = actual glob ownership, so a future reader doesn't read this event as a voice-room defect.
    • Cross-check the prior 3 lessons for 🎤 voice (fffc72f47, fe6d2be6e, fd408ad69) — all show offPct trending UP after facts-only interventions; confirm whether those were ALSO backend/track commits mistagged to voice, which would mean this is a recurring mistagging pattern, not three unrelated content-drift events.

    What we learned: This is the room-attribution version of THE PROXY IS NOT THE THING: the lens scores lane-fit against Originating-Terminal (which terminal happened to be open) rather than against which room's owned-surface glob the diff's files actually match — a commit trailer already carries the correct answer (Relevant-Rooms: 🔨 builder, 🧪 laboratory, 🔒 vault, explicitly NOT voice) but the offPct measurement ignores it and grades voice's lane-fit anyway. Three prior 'voice' interventions with rising offPct trends may be the same mistagging repeating, not three failures of the room's actual work.

    The missing guard (the deliverable): A structural test asserting: for any commit, if Relevant-Rooms omits the Originating-Terminal room, the lane-fit/offPct measurement must be scored against a Relevant-Room (file-glob match), never against Originating-Terminal alone — i.e. a commit whose own trailers already disclaim the originating room's ownership should not be graded as an in-lane failure for that room.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litIntent (126) and litReality (125) are nearly identical and walksIntent/walksReality (126/125) track them almost 1:1 — the walk is NOT under-filled, but walkTruncated=true flags it hit the depth/time budget before fully resolving, and with fill only 0.85% (well under the 70% ceiling) there's headroom to let it run one more ply rather than truncating, which might resolve whether the mass genuinely belongs at Tactics×Operations or would settle closer to a track/analytics-specific coordinate given more depth. Experiments: Re-run this exact commit with --max-depth+1 (one more ply) and diff the resulting offPct against the current 25 — if it drops meaningfully, walkTruncated was masking a better-fitting coordinate and the depth budget should be raised for track/API-glob commits specifically. · Seed the reef with src/app/api/track/**/*.ts as a labeled corpus fragment under whichever room's owned-surface eventually claims it, then re-run prewalkOverlap for this sha — if overlap jumps from 3 to something reef-recognized, that confirms the low overlap was a corpus-coverage gap, not a genuine cross-lane event. · Compute offPct for this same diff scored against 🔨 builder's owned surface instead of 🎤 voice's — if it comes back under the 15 threshold, that's direct evidence the event is a room-attribution bug (mistagged) rather than a real drift, closing the loop on the boundary_failure classification above.

    Measured outcome (deterministic, not self-graded): room mean offPct 23→15 over 6 commits since · dictionary ablation saved 589B

    An out-of-lane event driven by a mistagged room (not bad work) inflates 🎤 voice's apparent loss ratio without cause; premium is restored by fixing the attribution scorer to key off file-glob ownership, not terminal-of-origin, so voice's rate reflects only its actual content-lane drift.

    Open the signed attestation →

40. 2026-09-04 · 🔨 builder · offPct 31 · structural

    feat(initiative): the milestone is the pull — a positive goal, not only a fence

    Why it left the lane: The diff creates scripts/ops/initiative-pull.mjs, edits scripts/ops/initiative-frame.mjs, and adds tests/ops/initiative-pull.test.mjs — none of which match any glob in 🔨 builder's declared owned surface (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The vocabulary that dragged the mass — 'data/milestones-next.json', 'session-close', 'pullFor'/'pullBrief', 'framePrompt' — is operator-pipeline vocabulary, not builder UI/build-script vocabulary. Crucially, the commit's OWN trailer already names the pull correctly: 'Relevant-Rooms: 🔨 builder, 🎩 operator, 🧭 navigator', and per CLAUDE.md's own COMMIT MESSAGE STRUCTURE section that trailer is computed by 'the 3 highest-scoring matches against .workflow/rooms/*.html Owned file/dir surface globs' — meaning a SEPARATE, already-working computation in this repo agrees scripts/ops/* belongs to operator. The receipt pipeline's intent-lane just never consults it.

    The semantic pull: 'scripts/ops/initiative-pull.mjs', 'scripts/ops/initiative-frame.mjs', 'tests/ops/initiative-pull.test.mjs', 'data/milestones-next.json' — all under scripts/ops/ or tests/ops/, a coordinate the Relevant-Rooms trailer itself scores highest for 🎩 operator, not 🔨 builder.

    Intent–reality gap: The commit claims builder-lane work (it was authored from, and receipted against, the builder terminal) while the receipt measured 138 lit reality cells landing almost entirely in scripts/ops/ and tests/ops/ — the operator coordinate the commit's own Relevant-Rooms trailer had already named as more relevant than builder.

    • Locate the exact script that currently resolves a commit's intent-lane glob set for offPct (grep scripts/pmu/ for 'ownedSurface' or the room-glob loader) and confirm it reads Originating-Terminal only.
    • Wire that resolver to read Relevant-Rooms from the same commit-trailer computation (or recompute the same `.workflow/rooms/*.html` match) rather than reimplementing room ownership a second, divergent way.
    • Re-render sha 224f767a38's panel with the union-lane fix and confirm offPct drops toward the 15 threshold with no change to sigmaDrift/driftPct's own math.
    • Ablation-check the fix against the room's last 55 out-of-lane events: does the union rule explain the majority of them, or is a residual cluster genuinely intentional/mistagged cross-room work?
    • Add the reconciliation guard (below) in the same commit as the resolver fix.
    • Update CLAUDE.md's COMMIT MESSAGE STRUCTURE section with the spec_rewrite text above so the next session doesn't relearn this gap.

    What we learned: This repo computes 'which room owns this commit' in TWO independent places that can disagree on the same commit: the Relevant-Rooms trailer (glob-matched against `.workflow/rooms/*.html`, correct here — it named operator) and the receipt pipeline's offPct intent-lane (apparently Originating-Terminal-only, wrong here). Because CLAUDE.md itself sanctions terminal/file-relevance divergence as normal and expected, the receipt is structurally guaranteed to misfire on exactly the commits the trailer already flagged honestly — and at 55 out-of-lane events in 14 days for this one room, that's not noise, it's the dominant failure mode for 🔨 builder.

    The missing guard (the deliverable): tests/pmu-simulator/receipt-lane-matches-trailer.test.mjs — for a commit whose Relevant-Rooms trailer names a room OTHER than Originating-Terminal, render the receipt with the union-of-owned-surfaces lane and assert offPct stays ≤ threshold (SMOKE: this sha, reintroduced, must go from 31 to ≤15); and for a commit with Originating-Terminal==sole Relevant-Room, assert the union changes nothing (NO SMOKE: normal in-lane commits stay unaffected).

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Both sides are fully saturated (100% lit-to-walked ratio) yet walkTruncated=true — the walk is being cut off by its depth/step budget before it can determine whether ANY path exists back into the claimed lane, so a low-overlap commit like this one gets maximal offPct by construction rather than by genuine unreachability. Concrete adjustment: when prewalkOverlap falls below ~10, raise --max-depth (or the step budget) specifically for that commit's walk before reporting a final offPct, since the truncation is most punishing exactly where the reef signal is weakest. Experiments: Feed the Relevant-Rooms trailer's own glob union into the reef seed as the intent lane for this sha and re-render; measure whether offPct falls to ≤15 while offPct on a random sample of 20 already-in-lane commits stays within noise of its current value. · For commits with prewalkOverlap<10, double --max-depth and re-run; measure whether litReality/walksReality drops below 100% (finds true structure) versus staying saturated (confirms the walk is genuinely exhausting the space, not just truncating). · Back-test prewalkOverlap as a leading indicator: over the last 90 days of receipts, compute the correlation between prewalkOverlap<10 and offPct>15 to see if a pre-walk warning can fire before the expensive walk even runs.

    Measured outcome (deterministic, not self-graded): room mean offPct 19→14 over 3 commits since · dictionary ablation saved 1229B

    An out-of-lane event that traces to two disagreeing room-ownership computations — one of which (the trailer) already had the right answer — is a pricing-noise problem, not a behavior problem, so it inflates this lane's apparent loss ratio without reflecting real drift; premium is restored the moment the receipt's lane resolver reads the same Relevant-Rooms union the trailer already computes.

    Open the signed attestation →

41. 2026-09-04 · 🔨 builder · offPct 22 · mistagged

    feat(bookclub): the sticker says what tapping it does, and every iamfim door names itself

    Why it left the lane: Every changed path in this commit — scripts/bookclub/render-email.mjs and tests/comms/sticker-cta-is-attributable.test.mjs — matches zero globs in builder's declared owned surface (src/app/, src/components/, src/lib/, public/, build-book-html.sh, convert-chapters-to-html.mjs, hooks/pre-commit, voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The diff's own vocabulary is attribution/copy, not renderer-architecture: 'utm_content=${encodeURIComponent(slot)}', 'dest: slot ? `${url}#${slot}`', and the new reader-facing sentence 'Next time you wonder how to help — tap the sticker.' The intent lane was seeded from a glob list with nothing under it for this diff, so the walk's reality mass had no in-lane cell to settle into and spread toward the comms/tracking-adjacent region instead — the receipt is reporting a room-surface gap, not a bad edit.

    The semantic pull: Two clusters pulled the reality walk off builder's declared surface: the token/analytics cluster (slot, utm_content, signed dest, message_id, dispatch_id) which is network/attribution vocabulary, and the copywriting cluster ('names the action', 'above the masthead', 'tap the sticker', 'that is the whole thing') which is voice vocabulary. Both live inside scripts/bookclub/render-email.mjs, a file the room edits constantly (45 out-of-lane events/14d) but that appears in none of builder's globs.

    Intent–reality gap: The commit message and trailers claim builder-room renderer work ('Builder owns the renderer') dispatched from the builder terminal, and the Story trailer even names the split explicitly ('voice owns whether the sentence earns its place'). The receipt measured reality landing almost entirely outside builder's declared lane (offPct 22, litReality 136 vs litIntent 134) because the room's own glob list was never updated to include the file where this 'renderer' work actually lives.

    • Add scripts/bookclub/** and tests/comms/** to builder's owned-surface glob wherever it's declared (room-surface config referenced by this event's context block)
    • Extract track()/slot logic from scripts/bookclub/render-email.mjs into src/lib/comms/tracking.mjs
    • Update the commit-trailer convention doc: a Story naming split ownership must set Relevant-Rooms' lead entry to the glob-matching room, not Originating-Terminal
    • Add a mistag guard: on commit, warn if zero touched paths match Originating-Terminal room's owned-surface globs
    • Run recontext to mark scripts/bookclub/** and tests/comms/** as CONTESTED across builder/voice/network until the split is settled
    • Backfill the last 45 builder out-of-lane events and check what fraction are scripts/bookclub/** or tests/comms/** paths — confirms or kills this as the room's dominant offPct driver

    What we learned: This room's prior three lessons were facts-only (no sensemaking) and its offPct trend has only gotten worse under them (16→46, 15→16) — because the fix those interventions could reach for was code-level, and the actual defect here is upstream of the code: builder's declared owned-surface glob has never included scripts/bookclub/ or tests/comms/, a path this room edits on nearly every send cycle. A room can write good, tested code and still register as 22% out-of-lane forever if its own ownership declaration doesn't cover the files it actually maintains.

    The missing guard (the deliverable): A pre-commit or post-commit check that, for the Originating-Terminal room named in the trailer, tests whether ANY touched file matches that room's declared owned-surface globs — and fails loud (not silent) when the match count is zero, rather than letting the drift receipt discover it downstream 22 points out of lane.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Both sides hit litIntent/walksIntent = litReality/walksReality = 1.0 (every step lit a cell, nothing died unlit) while walkTruncated=true — so the walk is saturated but budget-capped, not naturally converged. For a mixed-domain commit like this one (renderer + attribution + copy in one file), the truncation may be masking whether depth would eventually pull reality back toward the intent lane or push it further out. Concrete adjustment: rerun this sha's receipt with --max-depth roughly doubled and compare offPct before concluding the walk is at fault versus the reef seed. Experiments: Rerun sha 29ce3c827's receipt with --max-depth 2x current and diff offPct against this run — isolates whether truncation, not seed placement, explains the residual spread · Pull prewalkOverlap across builder's last 45 out-of-lane events and correlate against offPct — if low overlap (<10) predicts high offPct across the sample, it confirms reef seed placement (not the walk) is this room's dominant driver and the fix belongs at the ingest/glob stage, not the walk config · Tag scripts/bookclub/** and tests/comms/** as a distinct 'comms' ingest class in the reef seed step (separate from whichever room commits it) and re-measure offPct for builder-authored bookclub commits under that seed class

    Measured outcome (deterministic, not self-graded): room mean offPct 19→18 over 12 commits since · dictionary ablation saved 902B

    22 points off-lane on a well-tested, deliberate fix reads as room-noise rather than defect-risk once the cause is a stale ownership glob, not bad code — premium for this lane should hold at current terms contingent on the glob patch landing, since an unpatched gap will keep manufacturing false-positive out-of-lane events on every future bookclub commit from this room.

    Open the signed attestation →

42. 2026-09-04 · 📐 architect · offPct 33 · mistagged

    chore(state): flush the day's derived ledgers and generated artifacts

    Why it left the lane: The diff's dominant semantic file is books/tesseract/chapters-html/04-you-are-the-proof.html, where a new `<!-- metavector:reach-is-verify-at-four-substrate-layers -->` blockquote was inserted carrying dense glossary cross-links (`/book/chapters/glossary#f4-verification-cost`, `#e2-fraud-detection`, `#e3-medical-ai`, `#g3-network`) and the note 'Refined for Book Club 2026-09-03 — Reach IS Verify at Four Substrate Layers.' That is hand-authored, glossary-linked book prose — a semantic profile close to the voice/book-craft coordinates, not the .thetacog/data ops-ledger coordinates the commit message describes. The reef's gzip-NCD seed for this commit (prewalkOverlap=5, very low against ~140-count walks) found little overlap between the declared lane's own ops-lexicon and the actual bytes staged, so the walk's mass drifted onto the book/glossary region.

    The semantic pull: `<!-- metavector:reach-is-verify-at-four-substrate-layers -->`, the four glossary hrefs, and the 220-line renumbered heading/paragraph block (headings shifting from data-line 301→315, 320→334, etc.) are real authored structural edits inside a chapter file — none of that vocabulary exists in the room's declared derived-artifact surface (.thetacog/*.ndjson, data/pmu/*.json, public/*panel*.png).

    Intent–reality gap: The commit claims '933 derived files — lens receipts, delegation/QC verdicts, mesh ledger, elicit pulse receipts, session-close state, book exports, per-commit panels. All recomputable, none hand-authored,' but the receipt measured hand-authored prose landing in the same commit: a new metavector annotation block with fresh glossary cross-references was written into books/tesseract/chapters-html/04-you-are-the-proof.html — the opposite of recomputable.

    • Split future 'flush the derived ledgers' commits from any diff touching books/**, src/content/**, docs/strategy/**, docs/comms/** — never bundle a state-flush with authored content.
    • Add the missing Persona-Intent trailer retroactively via `git notes add` on 2abd930e6 (per the amend-ban rule) naming the ideal reader of the reach-is-verify-at-four-substrate-layers metavector.
    • Note in .workflow/rooms/architect.html's owned-surface section that books/tesseract/chapters-html/** is explicitly NOT part of architect's lane.
    • Audit the room's other 176 out-of-lane events in the last 14 days for the same bundling pattern — a mass-flush commit is a recurring smuggling vector, not a one-off.
    • Add a derived-file allowlist manifest (e.g. `.thetacog/derived-file-globs.json`) that both the pre-commit hook and this detector's lane classifier can read, so 'recomputable' is asserted mechanically rather than typed by hand each time.

    What we learned: A 'flush the derived ledgers' commit is a magnet for silently smuggling hand-authored content, because at 933-files scale nobody re-diffs each path by eye — and the commit message's own factual claim ('none hand-authored') goes unverified at commit time. This is the inverse of the already-guarded 'state in a generated file' failure: here authored prose rides inside a commit claimed to be pure generated churn, and the volume itself is what hides it.

    The missing guard (the deliverable): A pre-commit check that, when a commit subject matches `^chore\(state\):` or the body asserts 'recomputable'/'none hand-authored', requires every staged path to match a maintained derived-file glob allowlist (.thetacog/**, data/pmu/**, public/*panel*/**, docs/pmu/commit-panels/**, etc.) — warn-only initially, listing any path (e.g. books/**, src/content/**) that falls outside it so the false claim is caught before commit rather than by the drift receipt after.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent ≈ 100% and litReality/walksReality ≈ 100% — the walk is fully saturated (every node it reaches also lights), which is too permissive to discriminate signal from noise on a 933-file mixed commit. Adjustment: lower --max-depth or tighten the lit-threshold for commits whose diffstat exceeds ~500 files, so saturation doesn't mask which subset of files actually drove the placement. Experiments: Seed the reef dictionary with the glossary-anchor href pattern (`/book/chapters/glossary#`) and re-run prewalkOverlap on this same commit; measure whether overlap rises meaningfully above 5. · Add a commit-type-aware seed corpus for chore(state)-typed commits (the known derived-file path list) and measure offPct across the room's next 10 chore(state) commits against the current baseline. · Re-run the receipt on two synthetic splits of this commit — derived-files-only vs book-HTML-only — to confirm the book file alone accounts for the full 33% off-lane mass rather than some smaller subset.

    Measured outcome (deterministic, not self-graded): room mean offPct 23→6 over 7 commits since · dictionary ablation saved 774B

    A flush commit whose own 'none hand-authored' claim is false erodes the architect lane's placement-claim reliability — each recurrence raises its loss-ratio by forcing manual reclassification of otherwise-trusted mass commits; premium restores once the allowlist guard ships and the room's next several chore(state) commits price back under the 15% lane threshold.

    Open the signed attestation →

43. 2026-09-04 · 🔨 builder · offPct 23 · mistagged

    docs(patents): record USPTO Track One classification milestone

    Why it left the lane: The diff itself is entirely `docs/patents/INDEX.md` (one status line) plus `tests/legal/patent-index-classification-status.test.mjs` (a 28-line guard). Neither string appears anywhere in 🔨 builder's owned surface (src/app, src/components, src/lib, public, scripts/build-book-html.sh, convert-chapters-to-html.mjs, hooks/pre-commit, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The reef has no lane at all for `docs/patents/` or `tests/legal/` under builder — every lit reality coordinate here is pulled by the terms 'patent', 'INDEX.md', 'classification', 'USPTO', 'Track One' toward legal/administrative ShortLex anchors, which sit structurally far from the UI/build-tooling lane builder claims. The commit message's own Relevant-Rooms trailer (🎩 operator, 📐 architect, 🔒 vault) confirms the author already knew builder wasn't the semantic home — it was never claimed as one, only as the Originating-Terminal.

    The semantic pull: 'docs/patents/INDEX.md', 'US 19/637,714', 'USPTO', 'Track One', 'classification' — none of these terms exist in builder's glob set (src/app, src/components, src/lib, public, scripts/build-book-html.sh, hooks/pre-commit, etc). The walk lit 120 reality cells while intent only covers 96, and prewalkOverlap is a modest 35 — the reef seeded weakly against builder's own lane and the walk's larger reality spread (120 vs 96) is exactly what you'd expect when the content has no true anchor in the claimed room and diffuses into whichever legal/ops coordinates 'patent'/'USPTO' actually resolve to.

    Intent–reality gap: The commit message frames this as builder-room 'internal reference-doc upkeep,' but the actual work — updating the master patent index and adding a legal-status guard — is operator/architect/vault territory by the commit's own Relevant-Rooms trailer. The room claim (Originating-Terminal: 🔨 builder) contradicts the content's real domain; the receipt is correctly reporting that the physical terminal a person happened to be sitting in when they typed this is not where the semantic mass lives.

    • Add owned-surface globs for docs/patents/** and tests/legal/** to whichever room's .workflow/rooms/*.html most plausibly owns legal/patent docs (📐 architect or 🔒 vault, per this commit's own Relevant-Rooms) so future patent-index commits have a real lane to land in
    • Audit .workflow/rooms/*.html for all 9 rooms to confirm none currently claims docs/patents/ or tests/legal/ — if true, this is a systemic gap, not a one-off
    • When the lane-assignment logic is touched, make it prefer Relevant-Rooms over Originating-Terminal for offPct scoring when the diff is 100% outside the terminal room's globs
    • Note in docs/patents/INDEX.md's own header (or nearby) that edits to this file are architect/vault-lane work regardless of which terminal the operator used, so a human skimming the file sees the convention too

    What we learned: The Originating-Terminal/Relevant-Rooms trailer convention already anticipates and correctly names cases where 'who typed it' and 'what it's about' diverge — this commit used the trailer exactly as designed — but the deterministic receipt pipeline that computes offPct doesn't appear to read Relevant-Rooms at all; it scores purely against Originating-Terminal's owned surface. The convention and the detector are two systems that were never wired together, so a commit can follow the documented convention to the letter and still generate an out-of-lane alarm.

    The missing guard (the deliverable): tests/pmu-simulator/relevant-rooms-informs-lane.test.mjs — asserts that when a commit's diff paths match zero globs in Originating-Terminal's owned surface but match globs in a named Relevant-Room, the receipt/lane-scoring path either (a) scores offPct against the Relevant-Room's lane instead, or (b) at minimum tags the event as 'terminal-vs-content mismatch, see Relevant-Rooms' rather than a bare threshold breach — so this class surfaces as a known, named case rather than 53 repeated raw alarms in one room's 14-day window.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Right-sized but poorly anchored: 96-120 lit cells on a 144-lattice is a healthy, non-saturated, non-truncated (walkTruncated=false) spread — the walk isn't shallow or capped. The problem isn't depth, it's that intent lane anchoring (from the commit message + builder's owned globs) gave the walk too few real intent seeds to converge against, letting reality diffuse wider. Adjustment: seed the intent side with Relevant-Rooms globs in addition to Originating-Terminal globs when computing walksIntent, so intent coverage isn't artificially starved on legitimately cross-room commits. Experiments: Re-run this exact commit's receipt with walksIntent seeded from the union of Originating-Terminal + Relevant-Rooms globs and measure whether offPct drops below the 15 threshold — if yes, this confirms Relevant-Rooms-aware seeding is the fix, not just a hypothesis · Backtest the room's other 52 out-of-lane events from the last 14 days against their own Relevant-Rooms trailers — measure what fraction are the same terminal-vs-content mismatch pattern (if >50%, this is the dominant drift class for 🔨 builder, not noise) · Compute prewalkOverlap distribution for commits where Originating-Terminal == the true semantic room vs commits where it doesn't, to calibrate a prewalkOverlap threshold (e.g. <40) that could flag mistagged-lane commits BEFORE the full walk runs, cheaper than post-hoc offPct

    Measured outcome (deterministic, not self-graded): room mean offPct 17→28 over 10 commits since · dictionary ablation saved 932B

    A single mistagged-lane event on administrative docs is low-severity noise, not a loss signal — but 53 out-of-lane events in 14 days in one room inflates that lane's apparent loss ratio with false positives, which erodes trust in the receipt itself; restoring premium requires the Relevant-Rooms-aware lane fix above so the detector stops charging builder for legal-domain content it never claimed.

    Open the signed attestation →

44. 2026-09-04 · 🔨 builder · offPct 37

    fix(initiative): the gate is a stencil, not a pinpoint — and the loop now produces a unit

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 17→27 over 12 commits since

    Open the signed attestation →

45. 2026-09-04 · 📐 architect · offPct 19 · mistagged

    feat(raise-deck): the meter seat stops being a toll booth — the log is what compounds

    Why it left the lane: The commit's two changed files — src/app/deck/raise/RaiseDeckClient.tsx and tests/deck/raise-deck.test.mjs — sit entirely outside 📐 architect's declared owned-surface globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md — every one a docs/HTML path, zero src/app or tests globs). The diff's load-bearing strings — 'DECK_VERSION', 'RaiseDeckClient', the meter-seat JSX block, the FABRICATED array reading docs/05-content/blog/scratchpad/BorisCap.txt — are all deck-engineering and test-authoring vocabulary, which the reef has no architect-owned lane to anchor to. With prewalkOverlap only 12 of 137 lit cells and walkTruncated=true, the walk saturated both sides fully (137=137=137=137) but started from a seed that already diverged, and the truncation gave it no budget to walk the mass back toward whatever intent cells docs/strategy/ does own.

    The semantic pull: The room's Story trailer says it plainly: 'this session takes the ARGUMENT into the raise deck' — the architect room is doing real strategy/argument work, but the argument now lives as JSX strings and prose comments inside a React client component and a test fixture reader, not inside any docs/ path the room's glob recognizes. Strings like 'the meter seat stops being a toll booth', 'the record compounds with it', and the whole BorisCap.txt fabrication-tripwire narrative are strategy-grade reasoning wearing code-file clothing.

    Intent–reality gap: The commit message and trailers claim this is 📐 architect work — strategy/argument reasoning — but the receipt measured the actual bytes landing in src/app/deck/raise/RaiseDeckClient.tsx and tests/deck/raise-deck.test.mjs, neither of which architect's owned-surface declaration recognizes as its terrain.

    • Add `src/app/deck/raise/**` and `tests/deck/**` to .workflow/rooms/architect.html's owned-surface glob list
    • Ship scripts/ops/room-surface-glob-check.mjs (warn-only pre-commit): resolve Originating-Terminal's owned globs against the diffstat paths, print the missing glob line when zero match — this is the SAME fix architect, navigator, and laboratory each independently spec'd per prior lesson ff09774e5 and none shipped; ship it now rather than re-spec it a fourth time
    • Re-run the boundary probe on 34849d766 with --max-depth raised past the current truncation point, to test whether offPct=19 is partly a walkTruncated artifact rather than a pure structural mismatch
    • Audit docs/strategy/ for whether raise-deck argument content (the moat sentence, the BorisCap register decision) should be authored there first and mirrored into RaiseDeckClient.tsx, matching the canonical RESEARCH→BOOK→BLOG source-of-truth pattern already used for book content
    • Cross-check whether docs/01-business/strategy/ should also gain a raise-deck pointer so a docs-only reader can find where the argument actually lives now

    What we learned: Prior lesson ff09774e5 diagnosed the DETECTION LOGIC (Originating-Terminal-only classification ignoring the sanctioned trailer divergence) and its fix (room-drift-<room>-check.sh) was independently spec'd three times but never shipped — this event shows the deeper problem is upstream of detection: the OWNERSHIP DATA itself (.workflow/rooms/architect.html's glob list) is stale relative to what the room actually produces. A room's Story trailer can correctly narrate new territory (the raise deck as 'the ARGUMENT') while its owned-surface declaration still only lists the old docs-only footprint — the detector isn't wrong about where the code landed, the room's own self-description is out of date.

    The missing guard (the deliverable): scripts/ops/room-surface-glob-check.mjs, run warn-only in pre-commit: for the commit's Originating-Terminal, resolve its .workflow/rooms/<room>.html owned-surface globs against the actual changed paths; if none match, print the room's file and the missing glob line to add. This is the guard three rooms already asked for and none shipped — the deliverable here is shipping it, not re-specifying it a fourth time.

    Detector probe: signal attribution — gzip reef 45% · ballistic walk 55%. Saturated, not shallow — litIntent/litReality both hit exactly 137/137, the walk's own cap, with walkTruncated=true flagging it explicitly. This is the 'saturated' failure mode from AR ledger, not under-use. Concrete adjustment: re-run this sha's receipt with --max-depth raised past the current ply ceiling and compare offPct; if it drops, truncation was inflating the reading and the glob fix alone may already be sufficient once re-measured. Experiments: Re-run the boundary probe on 34849d7665 with --max-depth increased beyond the current cap and diff the resulting offPct against 19 — isolates how much of the 19 is truncation artifact vs real seed divergence · After adding src/app/deck/raise/** to architect's owned glob, recompute this commit's intent lane against the new glob and re-render the panel to confirm offPct falls under 15 without any code change · Ablate the BorisCap.txt path string and the FABRICATED-array vocabulary from the reef corpus for this commit only, to measure how much of the walk's spread the fabrication-tripwire test content alone contributes versus the meter-seat copy change

    Measured outcome (deterministic, not self-graded): room mean offPct 11→27 over 7 commits since · dictionary ablation saved 919B

    An out-of-lane event driven by a stale ownership glob rather than bad work inflates this lane's apparent loss ratio for free — restoring it costs one config edit (the two globs) plus shipping the already-thrice-requested surface-check guard, not a code fix.

    Open the signed attestation →

46. 2026-09-04 · 🎤 voice · offPct 71 · structural

    fix(ops): guard-seen-to-fail journals its writes, because a finally does not survive SIGKILL

    Why it left the lane: The diff is 100% ops-infra vocabulary — scripts/ops/guard-seen-to-fail.mjs (INFLIGHT_REL, recoverInflight, SIGTERM/SIGKILL handlers, 'journal'), tests/ops/guard-seen-to-fail-survives-a-kill.test.mjs, docs/incidents/INCIDENT-REPORT-defect-bytes-survive-a-kill-2026-09-03.md — with zero terms from 🎤 voice's owned lane (books/tesseract chapters, blog mdx, decentralizedtalk, 6needs). prewalkOverlap=1 shows the reef seed itself already found near-zero shared mass between the intent lane (voice's book/blog glob) and the reality lane (ops vocabulary) before the walk even ran, which is why offPct lands at 71 with walksIntent/walksReality and litIntent/litReality staying almost symmetric (134/138) — the asymmetry isn't a walk artifact, it's baked into which glob got called 'intent'.

    Intent–reality gap: The commit's own Relevant-Rooms trailer already names 📐 architect, 🎩 operator, 🧪 laboratory — the commit itself asserts voice is NOT the relevant room. But scripts/interventions/intervention-fire.sh:73-75 resolves the receipt's 'room' solely from the Originating-Terminal trailer ('the room the operator was physically in'), so the receipt measured ops work against voice's content-glob lane the commit had already disclaimed.

    • scripts/interventions/intervention-fire.sh:74 — prefer Relevant-Rooms' first room over Originating-Terminal for ROOM resolution
    • Add scripts/ops/**, tests/ops/** as an owned-surface glob on .workflow/rooms/kitty-operator.html (🎩 operator / Operations grid) — currently unowned by any room's declared glob
    • Formalize docs/incidents/** as an owned glob on .workflow/rooms/wezterm-vault.html — currently only mentioned in prose, not a declared glob, per the same room's own text ('Any incident... gets an entry in docs/incidents/')
    • Close self-prompt sp-478 (due 2026-09-04) as resolved by this fix rather than letting it fire separately tomorrow — it already names this exact gap: 'scripts/digest/**, scripts/ops/** need ownership assertions'
    • Add a regression test asserting intervention-fire.sh resolves room from Relevant-Rooms when present, using this commit's own trailers as the fixture (Originating-Terminal=voice, Relevant-Rooms=architect/operator/laboratory → resolved room must not be voice)
    • Audit scripts/digest/** (named alongside scripts/ops/** in sp-478) for the same unowned-surface gap

    What we learned: The three prior interventions for this room (offPct trends 27→39, 0→36, 26→20 — two got WORSE) never touched the root cause because the root cause isn't behavioral, it's wiring: scripts/interventions/intervention-fire.sh:74 derives the receipt's 'room' exclusively from the Originating-Terminal trailer, which CLAUDE.md's own commit-structure section already documents as legitimately divergent from content relevance via the separate Relevant-Rooms trailer — and a self-prompt (sp-478, due tomorrow) had already independently discovered that scripts/ops/** and scripts/digest/** are unowned by any room's glob. This room's 98 out-of-lane events/14 days is very likely dominated by this one mechanical mismatch, not 98 separate content-relevance failures.

    The missing guard (the deliverable): A test asserting scripts/interventions/intervention-fire.sh resolves ROOM from the Relevant-Rooms trailer when present, falling back to Originating-Terminal only when absent — run against this exact commit's trailers (Originating-Terminal=🎤 voice, Relevant-Rooms=📐 architect, 🎩 operator, 🧪 laboratory) and asserting the resolved room is NOT voice.

    Detector probe: signal attribution — gzip reef 85% · ballistic walk 15%. Right-sized, not shallow or saturated — litIntent≈litReality with <3% spread difference means the walk is finding comparable mass on both sides; the defect is upstream of the walk, in which glob got called 'intent'. No walk-depth adjustment indicated. Experiments: Re-run this commit's receipt forcing intent lane = the union of architect+operator+laboratory owned globs (per Relevant-Rooms) instead of voice's glob; predict offPct drops well under 15 · Add scripts/ops/**, tests/ops/**, docs/incidents/** as an owned glob on one room and re-run the same commit against BOTH old and new room resolution to isolate how much of the 71 points was the room-choice vs the ownership-gap · Backtest all 98 out-of-lane events for 🎤 voice over the last 14 days, partition by whether Relevant-Rooms would have picked a different room than Originating-Terminal, and measure what fraction of the 98 resolves purely from the code_rewrite

    Measured outcome (deterministic, not self-graded): room mean offPct 23→19 over 4 commits since · dictionary ablation saved 974B

    This isn't real semantic drift, it's a chronic false-positive in the lane meter itself — at 98 events/14 days it's padding 🎤 voice's loss-ratio with noise that masks any genuine drift underneath, and the lane's premium is only restored once room-resolution reads Relevant-Rooms and scripts/ops/**·tests/ops/**·docs/incidents/** get a declared owner, so offPct on this room becomes a trustworthy signal again rather than a terminal-location artifact.

    Open the signed attestation →

47. 2026-09-04 · 🔨 builder · offPct 29 · structural

    fix(chat): the gmail redirect URI was never registered — discover it, never assume it

    Why it left the lane: None of the diff's real paths — packages/thetacog-mcp/gmail-sync.js, scripts/chat/sync-verdict.mjs, tests/.../gmail-auth-redirect-is-discovered.test.mjs — match any glob in builder's owned surface (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, .git/hooks/pre-commit, scripts/voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The receipt's room comes from Originating-Terminal (🔨 builder — where the operator physically sat) not from file relevance, and the commit's own Story trailer says this is network-delegated work: 'The network room delegated this after the network room delegated this after a live deal file read a real 2026-08-28 reply from Jason Bishara as silence.' Relevant-Rooms already lists 🔨 builder, ☕ network, 🎩 operator — the commit itself flags the divergence CLAUDE.md's own trailer convention calls 'fine and worth recording.'

    The semantic pull: gzip-NCD reef placement plausibly pulled toward builder because 'packages/thetacog-mcp/gmail-sync.js' textually rhymes with builder's owned 'thetacoach-crm-mcp/' glob — both are 'theta*-mcp' package names differing only in 'coach-crm' vs 'cog' — and 'scripts/chat/sync-verdict.mjs' sits in the same scripts/*.mjs neighborhood as builder's owned scripts/build-book-html.sh and scripts/voice-*-check.sh. prewalkOverlap=3 (tiny against litIntent=126/litReality=113) shows this seed match was thin even so — a rhyme, not a real hit.

    Intent–reality gap: The commit message and its own Story trailer are honest that this is 'the builder's half' of a fix delegated by the network room over a Gmail/chat-spine bug — the CLAIM is scoped correctly. The receipt measured 29% off-lane against builder's owned surface because the pipeline's lane is keyed to Originating-Terminal (physical terminal at commit time), and CLAUDE.md's own commit-trailer rule already states terminal and file-relevance 'can diverge — that's fine.' The gap isn't a bad claim, it's a receipt architecture that hasn't caught up to a documented, sanctioned pattern.

    • grep -rl "Originating-Terminal" scripts/pmu scripts/controller to find the exact lane-assignment script and patch it per code_rewrite
    • Add a delegation-aware owned-surface entry for packages/thetacog-mcp/** and scripts/chat/** to whichever room genuinely owns the message/chat spine (☕ network per THE MESSAGE SPINE memory), so future non-delegated commits there don't misfire onto builder
    • Apply the spec_rewrite text to CLAUDE.md's COMMIT MESSAGE STRUCTURE section in the same commit as the code fix
    • Audit builder's 54 out-of-lane events in the last 14 days for the same signature (Story trailer contains 'delegated' + Originating-Terminal: builder) — if most match, builder's headline offPct is systematically overstated, not actually drifting
    • Name the thetacog-mcp vs thetacoach-crm-mcp homonym explicitly in CLAUDE.md (same pattern as the pmu-rust/PMU homonym in AXIOM 0) so the reef's gzip-NCD stops rhyming the two package names

    What we learned: The drift-receipt's lane is keyed to Originating-Terminal (physical terminal), but CLAUDE.md's own commit convention explicitly licenses Originating-Terminal to diverge from content ownership on delegated work — so every cross-room delegation executed from the delegate's terminal is structurally guaranteed to score as out-of-lane, and a room that absorbs a lot of delegated work (builder: 54 events/14d) will show an inflated offPct that isn't measuring real drift at all.

    The missing guard (the deliverable): tests/pmu-simulator/delegated-commit-lane-resolution.test.mjs — SMOKE: a synthetic commit with Originating-Terminal: builder, a Story trailer containing 'delegated', and changed files matching only network's owned globs must resolve lane=network and score in-lane. NO SMOKE: an otherwise-identical commit with no delegation language in Story must still resolve lane=builder and is allowed to legitimately score off-lane if the files don't match.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized, not saturated or starved — litIntent/litReality (126/113) sit well under the 144-cell ceiling and are close in magnitude, so the walk isn't truncated (walkTruncated=false confirms) and isn't collapsing to a handful of cells either. Concrete adjustment: none dramatic needed; worth testing one extra --max-depth ply specifically on Story-trailer-delegated commits, since cross-room-boundary text may need an extra hop to reconverge intent and reality onto the same lane. Experiments: Strip the shared 'mcp' token from the reef's ownership tokenizer for room-seeding and re-measure prewalkOverlap on builder's known-delegated commits — expect it to rise, evidence the rhyme (not real content) drove the seed · Re-run this exact commit's receipt three times, once per Relevant-Rooms entry (builder/network/operator) as the forced lane, and compare offPct — if network/operator score dramatically lower, it confirms lane misassignment rather than real drift explains builder's 54-event/14d rate · Ablation-test seeding 'packages/thetacog-mcp' and 'thetacoach-crm-mcp' as distinguished (never-merged) tokens in the gzip dictionary, then measure whether builder's offPct on true thetacoach-crm-mcp commits stays flat while thetacog-mcp-only commits' pull toward builder drops

    Measured outcome (deterministic, not self-graded): room mean offPct 19→14 over 3 commits since · dictionary ablation saved 905B

    A structural mislane like this one dents builder's measured loss-ratio on paper without any real drift having occurred, which is worse for underwriting than an honest miss — it hides the signal inside noise; restoring it means the delegation-aware lane fix and guard ship together so offPct on delegated commits reflects actual placement, not terminal geography.

    Open the signed attestation →

48. 2026-09-04 · 🔨 builder · offPct 18 · mistagged

    feat(initiative): sufficiency on the INPUT — the restoring force that makes the loop converge

    Why it left the lane: The diff is entirely scripts/ops/initiative-admit.mjs, scripts/ops/initiative-mine.mjs, tests/ops/initiative-admit.test.mjs — vocabulary like 'band', 'calibrate', 'admit', 'MIN_GZIP_BYTES', 'spanMass', 'massOf()' and the scripts/ops/ + tests/ops/ paths themselves. Checked directly against .workflow/rooms/*.html: NO room's Owned-Surface glob claims scripts/ops/** or tests/ops/** — it is unclaimed territory, not contested territory. The reality mass had nowhere legitimate to land, so it registered as 18% off whatever lane the commit was scored against (builder's src/app · src/components · src/lib · public · book-build scripts · hooks/pre-commit · voice-check · decentralizedtalk · thetacoach-crm-mcp · vercel.json — none of which this diff touches at all).

    The semantic pull: Every changed path in the diff (scripts/ops/initiative-admit.mjs, scripts/ops/initiative-mine.mjs, tests/ops/initiative-admit.test.mjs) sits in a directory with zero owned-surface glob coverage anywhere in .workflow/rooms/*.html, confirmed by direct grep. 🧪 laboratory's own committed focus_areas field for today already reads '2026-09-04 — scripts/ops (57c, test-anchored) · initiative / iamfim / never' — the room itself has already noticed this exact territory is its live work and nobody has formalized it into the glob yet.

    Intent–reality gap: The commit message correctly self-describes as pipeline/ops infrastructure work (an input-sufficiency gate for the initiative-mining loop) and even lists Relevant-Rooms as 🔨 builder, 🧪 laboratory, 🎩 operator — but Originating-Terminal is tagged 🔨 builder, whose Owned Surface is UI/app/book-build/CRM territory the diff never touches. The commit already knew it was multi-domain; the single-lane offPct math scored it only against the terminal tag.

    • Ship scripts/ops/room-surface-glob-check.mjs per sp-483 (spec'd, not yet on disk — confirmed via `ls`) — warn-only pre-commit check that every staged path is claimed by ≥1 room's Owned Surface
    • Ship tests/ops/room-lane-trailer-union.test.mjs per sp-494 (spec'd, not yet on disk) — seeds a synthetic commit whose Originating-Terminal glob excludes the changed file but whose Relevant-Rooms trailer covers it, asserting the receipt should score against the union, not the single tag
    • Add scripts/ops/** + tests/ops/** to .workflow/rooms/cursor-laboratory.html's Owned Surface (or split with kitty-operator.html if laboratory/operator later contest it — run recontext to decide)
    • Replay commit 470752c93's receipt with the updated glob and confirm offPct drops under the 15 threshold — the ablation that proves the fix rather than assuming it
    • Backtest the other 59 out-of-lane 🔨 builder events from the last 14 days against the same glob change to measure what fraction share this exact unowned-scripts/ops/ root cause
    • sp-483 and sp-494 are due 2026-09-05 (tomorrow) — this event is the concrete evidence to close them; don't let the guard slip past its own due date

    What we learned: An entire, actively-growing infra directory — scripts/ops/ (rules-for.mjs, wip.mjs, quiet-gate.mjs, guard-seen-to-fail.mjs, token-vs-construct.mjs, initiative-admit.mjs, initiative-mine.mjs, and more) plus tests/ops/ — has no owner in any of the 9 rooms' .workflow/rooms/*.html glob lists, confirmed by direct grep rather than inferred. This is why 🔨 builder alone logged 60 out-of-lane events in 14 days: any commit authored from whichever terminal happens to be open when ops-tooling work gets done registers as off-lane, because there is no correct lane to register against. Prior facts-only interventions in this room measurably WORSENED the trend (16→46 after fa2c9b092) precisely because they had no sensemaking to point at this root cause — reactive re-scoring without diagnosis, exactly the SELF-IMPROVING FIRES failure mode CLAUDE.md already names.

    The missing guard (the deliverable): scripts/ops/room-surface-glob-check.mjs + tests/ops/room-lane-trailer-union.test.mjs — both already spec'd as self-prompts sp-483/sp-494 (due 2026-09-05) but neither exists on disk yet (confirmed via `ls`, both return 'No such file or directory'). The guard: every changed path in a commit must be claimed by at least one room's Owned-Surface glob, warn-only pre-commit; the test seeds the exact synthetic case this event demonstrates for real.

    Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. litIntent==walksIntent (108) and litReality==walksReality (140) — every walked node lit, so the walk is not under-firing. But walkTruncated=true means it hit its ply/depth budget before either side fully settled, and the reality side needed ~30% more capacity than intent to place initiative-admit.mjs's vocabulary (band, calibrate, spanMass, MIN_GZIP_BYTES, admit). Adjustment: raise --max-depth specifically when litReality/litIntent exceeds ~1.2x, since that ratio is itself the early-warning signal of a thin intent-side corpus rather than a genuine drift. Experiments: Once scripts/ops/room-surface-glob-check.mjs ships (sp-483), run it over `git ls-files` and confirm scripts/ops/** + tests/ops/** report UNOWNED today — establishes ground truth before the config fix lands · Replay 470752c93's receipt with a synthetic Owned-Surface glob adding scripts/ops/**+tests/ops/** to cursor-laboratory.html and confirm offPct drops under 15 — the ablation that proves rather than assumes the fix · Backtest the room's other 59 out-of-lane events from the last 14 days against the same synthetic glob change, measuring what fraction resolve — quantifies how much of builder's 60-event/14-day rate is this one root cause versus others

    Measured outcome (deterministic, not self-graded): room mean offPct 20→14 over 3 commits since · dictionary ablation saved 1011B

    This is an attribution miss, not a work-quality defect — it dents 🔨 builder's paper loss-ratio while the underlying ops work is sound and already claimed by 🧪 laboratory's own stated focus; premium on this lane is restored the instant scripts/ops/**+tests/ops/** get a real owner in the glob config rather than by builder producing less infrastructure.

    Open the signed attestation →

49. 2026-09-04 · 🔨 builder · offPct 22 · mistagged

    feat(initiative): six axes, all six over 80, and the bar shows its working

    Why it left the lane: The entire diff lives in scripts/ops/initiative-{dispatch,execute,frame,score}.mjs and tests/ops/initiative-score.test.mjs. None of those paths matches any glob in 🔨 builder's owned surface (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, .git/hooks/pre-commit, scripts/voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json) — the words the reef ingested from the new file (identifiers like eligibility, axes, calibration, deepen, pullStrength, chainRung, milestones, weights) are pipeline/scoring vocabulary that projects onto operations-shaped coordinates, not onto anything in builder's declared lane, which is exactly why 138 lit reality cells (vs 83 intent) drifted 22% outside it.

    The semantic pull: The new file's own docstring names the pull precisely: 'six eligibility axes... computed from the ARTIFACT rather than from a model's self-report', plus function names score(), deepen(), calibrate() and the LEDGER/WEIGHTS constants in .thetacog/initiatives/ — this is a scoring/gating instrument, structurally identical in shape to the room's OWN receipt pipeline, and gzip-NCD grips that shape wherever it has seen it before (operations/scoring coordinates), not wherever the committer's terminal happened to be.

    Intent–reality gap: The commit message claims a coherent unit of work ('every axis reads a structural property... both reach the email') and is trailer-tagged Originating-Terminal: 🔨 builder — but the receipt measures the work against builder's owned-surface glob, and 100% of the changed/new files (scripts/ops/*, tests/ops/*) sit outside it; the gap is between 'which terminal I was typing in' and 'which room's surface this code is'.

    • Add `scripts/ops/**` + `tests/ops/**` to whichever room's `.workflow/rooms/*.html` owned-surface actually claims general ops/pipeline scripts (likely kitty-operator.html or cursor-laboratory.html per the Relevant-Rooms trailer) — currently no room file in `.workflow/rooms/` visibly claims that glob.
    • Ship `scripts/ops/room-surface-glob-check.mjs` (warn-only pre-commit) exactly as already specced in sp-483.
    • Ship `tests/ops/room-lane-trailer-union.test.mjs` exactly as already specced in sp-494: seed a synthetic commit whose Originating-Terminal glob excludes the changed file but whose Relevant-Rooms glob includes it, assert the receipt scores against the union/best-match, not the terminal alone.
    • Change the receipt-scoring call site (wherever offPct is computed against 'this room's owned surface') to accept a fallback room list from the Relevant-Rooms trailer when Originating-Terminal's own glob-match is 0.
    • Audit the room's 59 out-of-lane events in the last 14 days for the same signature (Originating-Terminal glob-match == 0, files under scripts/ops/ or similarly unclaimed dirs) to size how much of the chronic drift is this one mechanism versus genuine builder-lane sprawl.

    What we learned: The three prior interventions for this room were 'facts-only' and all three measured WORSE offPct afterward (16→46, null→26, 15→16) — not because the diagnosis was unavailable, but because a diagnosis that only reaches a self-prompt queue never reaches the receipt pipeline: sp-483 and sp-494 already named this exact mechanism and guard on 2026-09-05 and are still 'due', unshipped, while commits keep landing. The structural fact this event reveals: a 'due' self-prompt is not a build — recurrence continues precisely because the fix has been correctly diagnosed and repeatedly re-diagnosed without ever becoming a commit.

    The missing guard (the deliverable): tests/ops/room-lane-trailer-union.test.mjs (already specced in sp-494, unshipped): seed a synthetic commit whose Originating-Terminal glob excludes the changed files but whose Relevant-Rooms glob includes them, and assert the lane-scoring path falls back to the Relevant-Rooms union rather than penalizing the commit against a terminal-of-convenience tag.

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. walkTruncated=true on a commit adding a new 316-line file + 166-line test (554 insertions across 5 files) suggests the walk hit its depth/time budget before fully resolving this larger-than-usual surface — likely too shallow for this commit class rather than saturated. Concrete adjustment: scale --max-depth (or the time budget) with diffstat size, e.g. bump depth for commits adding ≥250 new lines in a single new file, and re-render to see if litReality stabilizes below 138 once untruncated. Experiments: Re-render this commit's receipt with scripts/ops/** claimed by a room and measure whether offPct drops from 22 toward the 15 threshold — a direct ablation of the spec_rewrite. · Re-render with --max-depth increased (walkTruncated forced false) and compare litReality/offPct to isolate how much of the 22% is walk-truncation noise vs genuine seed misplacement. · Re-score the same commit against the union of Relevant-Rooms globs (🔨🎩🧪) instead of Originating-Terminal alone and measure the offPct delta — the exact experiment sp-494 already specs.

    Measured outcome (deterministic, not self-graded): room mean offPct 19→42 over 4 commits since · dictionary ablation saved 470B

    An out-of-lane event whose own reef already diagnosed the fix and queued it (sp-483/sp-494) two intervention-cycles ago without shipping reads as a control-deficiency, not a one-off drift, and raises this lane's loss-ratio until the guard actually lands rather than being re-diagnosed a fourth time.

    Open the signed attestation →

50. 2026-09-04 · 🎤 voice · offPct 21

    fix(comms): the dispatch lock now names its holder, and I am the incident

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 20→34 over 8 commits since

    Open the signed attestation →

51. 2026-09-04 · 📐 architect · offPct 19 · structural

    fix(comms): a bold LEAD PARAGRAPH is prose, not a spine header

    Why it left the lane: No encircled-region coordinate narrative is cached for 5133682477, so the positional claim has to be reconstructed from the owned-surface mismatch and the receipt's own numbers rather than named lattice cells. The diff's actual mass is `scripts/comms/canonical-template-check.mjs` (the `BOLD_LABEL_MAX` bound, the `bold[1].trim().length <= BOLD_LABEL_MAX` guard) and `tests/comms/spine-scan-headers-only.test.js` (`execFileSync('git', ['show', ...])`, `expect(r.code).toBe(0)`, `expect(r.out).not.toMatch(/out of order/)`). None of that vocabulary — regex bounds, Jest fixtures, `git show` fixture-loading — appears anywhere in architect's declared reef (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md), which is pure prose/strategy vocabulary. prewalkOverlap=2 against litIntent=102 confirms this directly: the gzip-NCD seed found almost no shared mass between architect's strategy reef and this commit's code+test corpus before the walk even ran, so the walk (litReality=124 vs litIntent=102) mostly amplified a divergence that was already baked in at seed time.

    The semantic pull: Runtime/test-shaped strings — `BOLD_LABEL_MAX`, `bold[1].trim().length`, `execFileSync`, `DEFECT_REV`/`DEFECT_PATH`, `expect(r.code).toBe(0)` — pulled the reality walk into the operations/tactics lane where scripts/ and tests/ actually live, while the commit's Originating-Terminal trailer anchors intent at architect's strategy-doc lane. The pull is structural, not semantic noise: it is the same kind of mismatch every scripts/comms fix from this room will produce until the lane surface changes.

    Intent–reality gap: The commit's own Story trailer asserts 'Architect owns the checker surface' — a legitimacy claim for exactly the files that were touched — but architect's registered owned-file-surface glob (`.workflow/rooms/architect.html`) contains zero code paths, only docs/strategy and docs/03-product. The receipt has no way to see the Story's claim; it can only see the glob, so a commit the operator considers correctly-scoped still measures as 19% out of lane.

    • Widen .workflow/rooms/architect.html's owned-surface glob to include scripts/comms/** (and confirm/extend builder's + operator's glob to cover tests/comms/** and the mail-dispatch surface respectively, per this commit's own Relevant-Rooms claim)
    • Unify the three independently-spec'd room-drift-<room>-check.sh drafts (architect/navigator/laboratory) into one canonical scripts/pmu/room-lane-classify.mjs and delete the duplicates
    • Switch the offPct computation to score against the Relevant-Rooms union rather than Originating-Terminal's glob alone
    • Add a staleness guard: flag when >1 room has specced the identical fix and none has shipped it after N days — this is the actual reason ff09774e5's diagnosis didn't close the gap
    • Reword or gate the commit Story convention so 'Architect owns X' is only assertable when X already appears in architect's registered glob, or is auto-flagged as a pending glob-widening PR

    What we learned: ff09774e5 diagnosed the Originating-Terminal-vs-Relevant-Rooms gap and even measured an improvement (offPct 20→14), but the improvement didn't hold (this event is back at 19) because the actual fix — the room-lane classifier rewrite — was independently specced identically by THREE rooms (architect, navigator, laboratory) and shipped by NONE of them. The reason nobody shipped it is structural, not accidental: fixing the classifier means editing scripts/pmu/*, which is itself out-of-lane for every docs-owning room that would attempt it — the bug and its own fix live in the same blind spot, so each room's attempt to fix it would generate a fresh out-of-lane event and nobody has an incentive to be the one who eats that hit.

    The missing guard (the deliverable): A guard that fails when two or more rooms' commit history contains near-identical spec language (e.g. gzip-NCD similarity above a threshold between commit-message bodies) proposing the same unshipped fix — surfacing duplicate-spec paralysis before it silently repeats a third time, and forcing one room to actually own the delegation.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litReality=124/144 ≈ 86% lit is near-saturation for the reality side; walksReality=124 tracking 1:1 with litReality suggests every walked cell lit, which is a saturation signal rather than a shallow-walk signal. Adjustment: cap `--max-depth` on the reality walk when litReality/144 crosses ~0.85 and compare offPct before/after to see if the extra spread past that point is signal or lattice-filling noise. Experiments: Recompute prewalkOverlap for this exact commit using a Relevant-Rooms-union reef (architect+builder+operator combined) instead of architect-only, and check whether overlap rises materially above 2 — directly tests the union-classifier hypothesis before building it · Seed architect's reef dictionary with scripts/comms/** + tests/comms/** vocabulary and re-render this commit's receipt; if offPct drops below 15 it confirms the glob-widening fix works and is a self-contained ablation · Pull architect's 179 out-of-lane events over the last 14 days and bucket by whether the touched paths fall inside ANY Relevant-Room's owned-surface glob vs none at all — quantifies what fraction of the room's chronic drift is this exact misclassification class vs genuine scope creep

    Measured outcome (deterministic, not self-graded): room mean offPct —→26 over 8 commits since · dictionary ablation saved 799B

    A chronic 179-event/14-day out-of-lane rate for one room reads as an unpriced systemic risk rather than isolated drift, and premium only comes back down once the owned-surface glob and the Relevant-Rooms-aware classifier both ship — until then every legitimate cross-room fix from architect keeps generating the same false loss signal.

    Open the signed attestation →

52. 2026-09-04 · 🎤 voice · offPct 19 · mistagged

    feat(iamfim): grade the send-hour burst — 344 first-hour views, 6 engaged

    Why it left the lane: The whole diff lives inside scripts/track/iamfim-visits.mjs and adds ops/analytics vocabulary — 's.burst = { views: 0, engaged: 0, human: 0 }', the literal grade strings 'T2-ENGAGED' and 'T3-HUMAN-80', 'sends.map(...burst:s.burst...)', and the --json burst payload. None of that touches books/tesseract/chapters, src/content/blog/*.mdx, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/, or scripts/voice-*-check.sh — the voice room's entire owned surface. The reef seed for 'voice' has almost nothing in it that resembles 'burst', 'grade', 'HUMAN-80', or 'delivery window', so the walk had nowhere in-lane to land and spread into the operations/measurement region of the lattice instead — which is exactly where litReality (135) exceeded litIntent (120).

    The semantic pull: Analytics/telemetry terms ('burst', 'grade', 'T2-ENGAGED', 'T3-HUMAN-80', 'matured', '--json') pulled the walk toward operations/measurement coordinates; the commit's own Relevant-Rooms trailer (🧭 navigator, 🎩 operator, 📐 architect) independently names that same region and conspicuously omits 🎤 voice — the commit is self-diagnosing its own lane.

    Intent–reality gap: The commit is filed under room 🎤 voice (Originating-Terminal), but 100% of the diff is inside scripts/track/iamfim-visits.mjs, an analytics/tracking script with zero overlap with voice's content surface — and the commit's own Relevant-Rooms line names navigator/operator/architect, not voice, so the trailer contradicts the lane it was graded against.

    • Add scripts/track/*.mjs (and sibling scripts/track/ helpers) to the 🎩 operator or 🧭 navigator owned-surface glob in .workflow/rooms/*.html — it is currently unowned by any of the 9 rooms' listed globs
    • Patch the lane-resolution function feeding the drift receipt to prefer Relevant-Rooms/best-fit-glob over raw Originating-Terminal when they diverge
    • Audit the room's other 101 out-of-lane events from the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms mismatch pattern before assuming this fix is complete
    • Add the missing_guard test (below) in the same commit as the lane-resolution patch
    • Update the room dispatch label printed in the receipt row to show BOTH Originating-Terminal and Relevant-Rooms so a human reading the receipt can see the divergence directly, not just infer it from the trailer

    What we learned: CLAUDE.md explicitly sanctions Originating-Terminal diverging from file-relevance ('that's fine and worth recording') — but the receipt pipeline currently grades drift AS IF Originating-Terminal were the intent-lane claim, so it manufactures out-of-lane events for exactly the pattern the trailer rule was written to permit. The 102-events-in-14-days count for this room is likely dominated by this one structural conflation, not by 102 separate genuine scope violations.

    The missing guard (the deliverable): tests/ops/receipt-lane-prefers-relevant-rooms.test.mjs — SMOKE: feed the receipt generator a commit whose Originating-Terminal and Relevant-Rooms trailers name different rooms, assert the receipt's graded lane matches Relevant-Rooms (or the best-fit glob owner), not the raw Originating-Terminal string. NO-SMOKE: feed a commit where both trailers agree, assert the lane is unchanged from today's behavior.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent(120)/litReality(135) are close in magnitude and walkTruncated=true fired anyway on a 16-line diff — worth checking whether walkTruncated is tripping on a default depth budget that's too shallow for small commits generally, rather than reflecting genuine unresolved spread here. Adjustment: re-run this sha with --max-depth raised one tier and confirm litReality/litIntent converge further before concluding the walk itself needs tuning. Experiments: Recompute prewalkOverlap for this diff's vocabulary ('burst','grade','T2-ENGAGED','T3-HUMAN-80') against all 9 room reefs and confirm voice scores lowest — if so, add a pre-commit warn when prewalkOverlap ≤ 3 for the claimed room · Ablation: add scripts/track/*.mjs vocabulary to the operator (or navigator) reef dictionary, re-render this same sha under that room's lane, and measure whether offPct drops materially — confirms the better-fit room before committing the glob change · Sample 10 of the room's other 101 out-of-lane events from the last 14 days and classify each as Originating-Terminal/Relevant-Rooms-divergence vs genuine scope creep, to size how much of the 102 count this single fix actually resolves

    Measured outcome (deterministic, not self-graded): room mean offPct 23→35 over 6 commits since · dictionary ablation saved 765B

    A mistagged lane inflates this room's measured loss ratio with false positives, which under-prices genuine voice-room drift elsewhere in the same 14-day window; restoring the premium means fixing lane-resolution to use Relevant-Rooms, not patching this one commit's tag by hand.

    Open the signed attestation →

53. 2026-09-04 · 🔨 builder · offPct 17 · mistagged

    fix(gemini): the batch harvest drives Chrome over CDP — the Safari path is unreachable

    Why it left the lane: Every changed path in this commit — scripts/chat/gemini-web-all.sh, scripts/chat/sync-all.sh, tests/chat/gemini-harvest-uses-cdp.test.mjs, .claude/skills/inbox/SKILL.md — sits outside all six of builder's owned-surface globs (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, scripts/voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). I checked every room's 'Owned file/dir surface' list directly (.workflow/rooms/*.html) and scripts/chat/ is not claimed by ANY of the 9 rooms — not builder, not navigator either, despite the commit's own Story line naming navigator as the corpus owner. The diff's dominant vocabulary — 'gemini-cdp-scrape.mjs', 'Chrome over CDP', 'chrome-gemini profile', 'Apple events', 'ensureChrome', 'History.db' — is browser-automation/ops vocabulary with no home in the reef's room lattice at all, so the walk had nothing legitimate to anchor to and the mass scattered.

    The semantic pull: No cached encircled-region narrative exists for this sha, so I ground the pull in the owned-surface facts directly rather than invent coordinates: 100% of the diff's 4 touched files are chat-pipeline files, 0% are builder's app/component/build-pipeline files, and the reef has no room claiming scripts/chat/ or tests/chat/ to pull the mass toward instead — it's not that another lane won the tug-of-war, it's that no lane was contesting it and the receipt fell back to whatever the Originating-Terminal trailer said (builder, because iTerm2 was the physical terminal open).

    Intent–reality gap: The commit message correctly describes a browser-automation correctness fix (Chrome-over-CDP replacing Safari/Apple-events) and even names navigator as the room that 'owns whether the corpus stays current' in its own Story trailer — but the commit's Originating-Terminal trailer says builder, and the receipt scored the work against builder's app/build-pipeline lane, where none of the 4 changed files live.

    • Add scripts/chat/**, tests/chat/**, .claude/skills/inbox/** to navigator's 'Owned file/dir surface' list in .workflow/rooms/rio-navigator.html — verified absent from all 9 room files, not just misassigned to builder
    • Cross-check scripts/comms/**, tests/comms/** for the same unowned-surface gap (verified absent from all 9 rooms in this pass) and register them to whichever room the comms-dispatch pipeline actually belongs to
    • Re-run this commit's receipt after the glob registration lands, to measure whether offPct drops once scripts/chat/ has a home lane
    • Audit builder's other 45 out-of-lane events from the last 14 days for the same path-outside-all-globs signature — if a majority are unowned-surface rather than genuinely builder-mistagged, the room's rising offPct trend (16→46 across the last 3 lessons) is a registration gap, not a discipline problem
    • Add a line to CLAUDE.md's SOVEREIGN OUTREACH / room-surface section cross-referencing which room owns scripts/chat/ once decided, so a future session doesn't have to re-discover the gap by grep

    What we learned: scripts/chat/ (and, checked in the same pass, scripts/comms/ and tests/chat/comms/) are UNOWNED surface across all 9 rooms' declared globs — not misassigned to the wrong room, genuinely unclaimed by any. Originating-Terminal is detected from the physical terminal app (open-room-session.sh --print-label), which is decoupled from which files a commit actually touches; when work lands on unowned surface, the trailer silently defaults to whichever terminal happened to be open rather than surfacing the gap. The three prior lessons for this room were all 'sensemaking unavailable, facts-only,' and its offPct trend has been worsening (null→26→46) across them — this is the first pass to find that a chunk of that trend may be structural (unowned surface) rather than behavioral drift in builder's actual lane.

    The missing guard (the deliverable): A commit-time check that cross-references every changed path against every room's owned-surface globs in .workflow/rooms/*.html and, when NO room's glob covers >0% of the changed paths, blocks silent Originating-Terminal defaulting and instead prints 'UNOWNED SURFACE: <paths> — assign before committing' rather than letting the physical-terminal detector pick a room by default.

    Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. Right-sized in spread (127-129 lit cells, near-symmetric) but walkTruncated=true flags it stopped at a budget rather than converging — for a 4-file, ~111-line diff, 127+ lit cells on each side is a lot of surface for the walk to still be truncating on. Concrete adjustment: raise --max-depth (or the ply time budget) for commits under a small line-count threshold, since a small diff hitting truncation before convergence suggests the budget is tuned for larger commits and is cutting off small ones mid-walk rather than letting them settle. Experiments: Register scripts/chat/**, tests/chat/**, .claude/skills/inbox/** to navigator's owned surface, then re-render this sha's receipt and measure whether offPct drops from 17 toward the 15 threshold — direct ablation of the unowned-surface hypothesis · Run guard-seen-to-fail against a synthetic room-ownership guard: verify it goes red replaying this commit's pre-fix trailer state (builder claiming scripts/chat/ work) and green after the glob registration + correct trailer · Sweep all 9 room HTML files' Owned file/dir surface lists against the full scripts/**, tests/** directory tree and report every top-level directory with zero owning room — scripts/chat/ and scripts/comms/ are two found in this pass; there are likely more

    Measured outcome (deterministic, not self-graded): room mean offPct 18→14 over 8 commits since · dictionary ablation saved 735B

    An unowned-surface commit reads as an out-of-lane loss event even when the code is correct, inflating builder's apparent loss ratio for work that was never really builder's risk to carry — premium on this lane restores once scripts/chat/ has a registered owner and the receipt measures against the right room.

    Open the signed attestation →

54. 2026-09-04 · 📐 architect · offPct 48 · mistagged

    feat(ops): ask the inverse question — which claimed guards has nobody ever asked?

    Why it left the lane: The diff's actual mass sits entirely in scripts/ops/guard-seen-to-fail.mjs and tests/ops/guard-roster.test.mjs — functions like claimedGuards(), roster(), the CLAIM regex, and the --roster CLI branch. None of those paths fall inside architect's owned globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The commit message talks ABOUT CLAUDE.md at length ('27 claimed guards', 'Guarded by X') but never edits CLAUDE.md itself — the prose references architect's territory while the bytes land in ops/testing territory, which is exactly the shape offPct=48 measures.

    The semantic pull: Heavy repeated vocabulary from CLAUDE.md's own rule-writing register — 'HARD RULE', 'Guarded by', 'guard the class', 'proven/asked/never-asked' — pulls the reef toward architect's rule-authoring coordinate even though the commit's actual work is a CLI tool and its test, both filed under scripts/ops and tests/ops. The commit is dense with CLAUDE.md-shaped prose describing tooling that lives outside CLAUDE.md.

    Intent–reality gap: The commit message reads as an ops-tooling ship (a new --roster flag + its test), correctly scoped and well-fenced — but it is filed under Originating-Terminal 📐 architect, whose owned surface is strategy docs and CLAUDE.md, not scripts/ops. The receipt measured where the bytes actually went (scripts/ops, tests/ops) against architect's lane and found 48% of the lit mass sitting outside it.

    • Ship ONE canonical scripts/ops/room-lane-classify.mjs unifying the three independently-specced room-drift-<room>-check.sh fixes (architect/navigator/laboratory all spec'd it, none shipped) — classify intent lane from Relevant-Rooms trailer + file-glob ownership, not Originating-Terminal alone
    • Add an explicit owned-surface glob for scripts/ops/** and tests/ops/** to whichever room's .workflow/rooms/*.html actually claims ops tooling (Relevant-Rooms names 🎩 operator and 🔒 vault — pick one, name it)
    • Add 'Guarded by tests/ops/guard-roster.test.mjs' as its own CLAUDE.md claim so guard-seen-to-fail.mjs's roster does not blind-spot its own guard
    • Fold tests/ops/claimed-guards-exist.test.mjs's duplicated CLAIM regex onto the now-exported canonical one, as its own commit (already named as debt in this commit's message)
    • Run --roster against the fix once shipped and confirm CLAUDE.md's guard-roster.test.mjs itself moves from 'never-asked' to 'proven'

    What we learned: The tool shipped in this commit audits every guard CLAUDE.md claims by name — but the drift-lane receipt pipeline that produced THIS offPct=48 report is not itself named 'Guarded by' anything in CLAUDE.md, so it is structurally invisible to the very roster the commit just built. An auditor that cannot see itself is a new instance of the class this commit exists to catch.

    The missing guard (the deliverable): A room-surface-coverage test asserting every top-level directory under scripts/ and tests/ that receives commits is claimed by at least one room's owned-surface glob in .workflow/rooms/*.html — so an unclaimed directory (like scripts/ops/ and tests/ops/ appear to be today) fails loudly instead of silently misattributing every commit that touches it to whichever room happened to be the terminal.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. walkTruncated=true and litIntent≈walksIntent, litReality≈walksReality (near-100% conversion from walked to lit) together mean the walk saturated its budget rather than exploring broadly — it filled whatever depth it was given and stopped, truncated before finding more structure. Adjustment: raise --max-depth for CLAUDE.md-vocabulary-dense ops commits like this one and re-check whether litIntent/litReality separate further (revealing more structure) or converge (confirming the walk was already sufficient). Experiments: Seed scripts/ops/ and tests/ops/ path-prefixes explicitly into the reef coordinate dictionary, then re-ingest this same commit and measure whether prewalkOverlap rises above 4 · Recompute this commit's lane classification using the union of all three Relevant-Rooms globs (architect+vault+operator) instead of Originating-Terminal alone, and measure whether offPct drops below the 15 threshold · Re-run the pipeline on this sha with --max-depth raised one increment and diff litIntent/litReality/offPct against the walkTruncated=true baseline to see if truncation is masking or explaining the drift

    Measured outcome (deterministic, not self-graded): room mean offPct 20→5 over 10 commits since · dictionary ablation saved 881B

    175 out-of-lane events in 14 days makes architect's offPct signal noisy enough that a real strategy-lane drift and a correctly-placed cross-cutting ops commit read identically, which is what erodes this lane's premium; restoring it requires the unified lane classifier (file-glob + Relevant-Rooms) so legitimate ops work stops pricing as architect drift.

    Open the signed attestation →

55. 2026-09-04 · 🔨 builder · offPct 42

    fix(crewbar): the hiring bar is not chrome — gate it on the path, not the domain

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→18 over 4 commits since

    Open the signed attestation →

56. 2026-09-04 · 🎤 voice · offPct 18 · structural

    fix(book): the apostrophe was surviving as "39" in 154 anchors, and it held the send

    Why it left the lane: The claimed lane is books/tesseract/chapters/*.md prose content (room 🎤 voice owns book text), but the actual diff is almost entirely inside scripts/convert-chapters-to-html.mjs — the slug-generation function, entity-decoding logic, and a triple-slugify compatibility shim (slugify/legacySlugify/pickerSlugify) plus a new test file tests/book/anchor-entities-not-poisoned.test.mjs. Strings like 'function slugify', 'decodeEntities', 'legacySlugify', 'pickerSlugify', and the renderer object's heading()/paragraph() methods are build-pipeline/tooling code, not book-chapter markdown. That is exactly the kind of mass that lands on 📐 architect / 🔨 builder coordinates (build scripts, HTML generation, test infra) rather than the voice room's owned content lane — hence 18% of lit reality cells fall outside intent.

    The semantic pull: The commit message is voice-flavored ('the apostrophe was surviving', 'Goodhart's Law', book anchors, book-club send) which correctly ANCHORS the intent lane to books/voice content, but the work itself is a converter-script bugfix plus a new regression test. Terms like 'slugify', 'decodeEntities', 'renderer', 'heading()', and the entire diff being isolated to scripts/convert-chapters-to-html.mjs and tests/book/ drag the reality mass toward the tooling/build coordinate — the walk lit up build-script and test-infra territory because that's where every changed line actually sits, even though the narrative is 100% about book anchors.

    Intent–reality gap: The commit message and Story trailer frame this as book-content work (anchors, chapters, the book-club send), correctly tagged 🎤 voice — but the receipt measures where code landed, and 100% of the diff lines are in a build script + a new test file, which is architecture/build territory the voice room doesn't own.

    • Add scripts/convert-chapters-to-html.mjs and tests/book/*.test.mjs to 📐 architect's (or 🔨 builder's) owned-surface glob list, not 🎤 voice's, in .workflow/rooms/*.html
    • Update CLAUDE.md's room table to explicitly separate 'book CONTENT' (voice) from 'book BUILD PIPELINE' (architect/builder)
    • Cross-link scripts/bookclub/pick-passage.mjs's slug() and scripts/convert-chapters-to-html.mjs's pickerSlugify() with a comment pointing at each other, since the test file already notes they must stay byte-identical — a future edit to one without the other is the next version of this exact defect
    • Add a one-line note atop scripts/convert-chapters-to-html.mjs's slugify() pointing future editors at CLAUDE.md's BOOK DEEP-LINK ANCHORS section so the anti-regression protocol is discoverable in the file that most needs it

    What we learned: A commit's NARRATIVE room (what content it's ABOUT) and its CODE-SURFACE room (where the changed lines physically live) can diverge even when the commit message is entirely accurate and well-written — this repo's room-ownership globs key off file path, not story, so a pipeline bugfix discovered via book content will always mistag as voice unless the owned-surface globs for scripts/*.mjs and tests/book/* are explicitly assigned to architect/builder ahead of time.

    The missing guard (the deliverable): A pre-commit check (or the existing owned-surface glob config) that flags when a commit's Originating-Terminal room owns none of the changed files but Relevant-Rooms does — i.e., cross-check `git diff --name-only` against `.workflow/rooms/*.html` globs and warn when the committing room's own glob-match count is zero while a listed Relevant-Room's match count is nonzero, so mistagged-room drift is visible in the commit itself rather than only in the drift receipt hours later.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. walksIntent equals walksReality equals litIntent equals litReality (133/133/133/133) with walkTruncated=true — the walk saturated its ply/time budget symmetrically on both sides rather than exhausting the graph, which can mask real spread differences (a truncated walk that hit the same cap on both sides looks balanced even if reality's untruncated spread would be wider). Concrete adjustment: re-run this commit's walk with --max-depth raised one ply specifically for the reality side to check whether litReality grows past 133 once untruncated — if it does, the current offPct=18 is a floor, not the true measure. Experiments: Re-run prewalkOverlap computation with the book-build-pipeline vocabulary (slugify, decodeEntities, anchor, renderer) added to the architect/builder reef seed corpus, then recompute this commit's offPct — if overlap rises and offPct drops, the reef's term coverage for build-tooling work is the actual gap, not the walk · Ablate walkTruncated by doubling --max-depth for this sha only and diff litReality against the current 133 to quantify how much of the 18% off-lane mass is truncation artifact vs real drift · Compare offPct across the room's last 104 out-of-lane events split by whether the diff touches scripts/*.mjs vs books/*.md — if scripts-touching voice commits systematically show higher offPct, that confirms the mistagged-room hypothesis structurally rather than per-incident

    Measured outcome (deterministic, not self-graded): room mean offPct 24→28 over 4 commits since · dictionary ablation saved 1032B

    An out-of-lane build-pipeline fix correctly narrated as book content is a low-severity mistagged-room event that dents 🎤 voice's loss-ratio on paper without representing real cross-lane risk-taking; premium is restored by moving the owned-surface glob for scripts/convert-chapters-to-html.mjs and tests/book/* to architect/builder so the next identical fix prices correctly against the room that actually owns it.

    Open the signed attestation →

57. 2026-09-04 · 🎤 voice · offPct 44 · mistagged

    comms(bookclub): a dead <a> never ships either — link liveness beside panel liveness

    Why it left the lane: Every changed path is bookclub send-pipeline infra — scripts/bookclub/link-liveness.mjs (new), scripts/bookclub/send-daily.mjs (the preflight block), tests/ops/bookclub-link-liveness.test.mjs — and the vocabulary is pure ops: execFileSync('curl',...), urlStatus, ownHostTargets(), process.exit(broken.length?1:0), '--json'. None of it touches books/tesseract/chapters, src/content/blog, decentralizedtalk/slideData.ts or src/app/6needs — the four globs voice actually owns (commit-page, blog, pixel per terminal-voice.html). The reef walk correctly followed the diff's own words into Operations-flavored lattice cells (send pipeline, preflight, exit-code gating) and away from voice's content-authoring cells, which is exactly what a 44% offPct measures.

    The semantic pull: 'send-daily.mjs', 'link-liveness.mjs', 'curl','-sS','-o','/dev/null','-w','%{http_code}', 'own-host', 'UNREACHABLE', 'process.exit(broken.length ? 1 : 0)' — these are the strings a builder/operations reef seed grips hardest; they are the same idiom the diff's own comment names ('Same idiom as send-daily's urlStatus'), i.e. this file is admittedly one more instance of an existing ops pattern living outside voice's four owned surfaces.

    Intent–reality gap: The commit message never claims to be voice-room content work — it is an honest, well-documented preflight fix for the bookclub send pipeline, and its own Relevant-Rooms trailer already names builder and operator as more relevant than voice. The receipt nonetheless scores it against voice's lane (Originating-Terminal), which is the terminal the operator happened to be sitting in, not the room whose owned surface the diff touches — so a correctly-scoped, correctly-labeled commit reads as a 44%-out-of-lane event for a room it barely relates to.

    • CLAUDE.md COMMIT MESSAGE STRUCTURE — add the glob-over-terminal clarification (spec_rewrite above) in the same commit as the code fix
    • scripts/pmu/pipeline.mjs (or the actual lane-seed call site) — implement glob-match-first lane resolution before falling back to Originating-Terminal
    • Audit the room's other 103 out-of-lane events in the last 14 days for the same signature (scripts/*, tests/* paths with an Originating-Terminal that doesn't own them) to size how much of voice's climbing offPct (27→39, 0→36, 26→20 across three prior facts-only passes) is this exact mistagging class rather than genuine content drift
    • Add a guard asserting: for a commit whose Originating-Terminal room's owned globs match zero changed files while another room's globs match all of them, the receipt's lane seed equals the matching room, not the terminal
    • Confirm messages-network.html's scripts/bookclub/ claim is current (it is, dated 2026-08-24) so the fix has a real target room rather than guessing

    What we learned: This room's chronic out-of-lane rate (103 events/14 days, rising across three prior facts-only interventions with no sensemaking) is not scattered content drift — it is a fixed structural artifact: the receipt attributes lane-fit to whichever physical terminal was open, but this repo's own commit-trailer convention (and the operator's actual workflow) explicitly permits doing cross-room ops work from any terminal. Every bookclub-infra commit typed from the terminal-voice session will keep scoring as voice drift until lane resolution is decoupled from Originating-Terminal — no amount of writing better voice-room content fixes a metric that's measuring the wrong room.

    The missing guard (the deliverable): A guard (e.g. tests/pmu-simulator/lane-follows-glob-not-terminal.test.mjs) that takes a synthetic commit whose changed files match ZERO of Originating-Terminal's owned globs but ALL of another room's owned globs, and asserts the receipt's lane-fit computation seeds against the glob-matching room — with this exact commit (6adcb7288, 100% under the network room's scripts/bookclub/ claim, 0% under voice's commit-page/blog/pixel) as the real-vocabulary fixture per GUARD-SEEN-TO-FAIL.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. litIntent=133 and litReality=131 against a 144-cell lattice with walkTruncated=true means both walks are near-saturating the full lattice — too deep/wide for a commit this narrow (3 files, one clear concern), and saturation erodes the walk's discriminative power exactly when discrimination is what's needed to catch a mistagged room. Adjustment: for commit-scope walks under a small file-count threshold, reduce --max-depth (or tighten ply decay) so intent/reality stay well under full saturation, preserving separation instead of both sides converging toward 'lit everywhere'. Experiments: Re-run this commit's reef seed using the network room's owned-glob corpus (data/bookclub/, scripts/bookclub/, docs/comms/bookclub-*) instead of voice's as the intent lane, and measure whether offPct collapses toward the room's baseline — this directly tests the mistagged-lane hypothesis · Lower --max-depth for commit-scope walks under a 5-file change threshold and re-render this same sha's triptych, checking whether litIntent/litReality drop below the apparent ~133/144 saturation point while the lane verdict stays stable · Sample 10 of the room's other 103 recent out-of-lane events, re-seed each against a glob-match-resolved room instead of Originating-Terminal, and measure what fraction of the 14-day offPct trend (27→39→0→36→26→20) is explained by this exact mistagging class vs genuine content drift

    Measured outcome (deterministic, not self-graded): room mean offPct 24→28 over 4 commits since · dictionary ablation saved 1090B

    An out-of-lane receipt this large, repeated 103 times in 14 days under one room, reads to an underwriter as a room with no reliable boundary — even though every individual instance here is a correctly-scoped, well-tested commit filed under the wrong terminal; premium is restored not by writing tighter voice-room content but by fixing lane attribution so the loss ratio reflects actual content drift instead of terminal-location noise.

    Open the signed attestation →

58. 2026-09-04 · 🎤 voice · offPct 25 · structural

    comms(external): correct the draft — the address is known, and deliberately not here

    Why it left the lane: The event's room label (🎤 voice) comes from the Originating-Terminal trailer, but the diff touches only docs/comms/external/drafts/2026-09-03-bobby-gupta-the-record-you-did-not-author.md — a file that matches none of voice's owned-surface globs (books/tesseract/chapters/*.md, src/content/blog/*.mdx, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/). The diff's actual vocabulary — 'milestone email', 'investor memo marked confidential', 'audience field', 'arming time', 'broker channel', 'dual send' — is comms-pipeline language, not book/blog voice language. Tellingly, the commit's own Relevant-Rooms trailer already names ☕ network, 🎩 operator, 🔒 vault and pointedly does NOT name 🎤 voice — the commit itself is on record disagreeing with the room the receipt attributed it to.

    The semantic pull: No region narrative is cached for this sha, so I can't cite specific out-of-lane coords directly — but the mismatch is legible from the owned-surface diff alone: every load-bearing noun phrase in the changed text ('audience field', 'arming time', 'confidential… memo', 'broker channel is coordinated') belongs to the comms/dispatch domain the Relevant-Rooms trailer names, while zero terms match voice's book/blog/6needs vocabulary. The pull toward an out-of-lane placement is structural (wrong file domain), not a stray keyword collision within an otherwise-correct lane.

    Intent–reality gap: The receipt claims this commit belongs to 🎤 voice's lane (via Originating-Terminal), but the reality it measured is a comms-draft correction that the commit's own Relevant-Rooms trailer assigns to network/operator/vault. CLAUDE.md's own commit-structure rule already states 'a commit's originating terminal and its file-relevance to rooms can diverge — that's fine and worth recording' — meaning this divergence was anticipated as normal, yet the drift pipeline appears to gate offPct on the terminal-of-origin rather than the glob-ownership attribution the same document says `thetacog_overview` uses.

    • Grep the post-commit/receipt pipeline for where it reads 'Originating-Terminal:' and confirm it feeds the room field used for offPct computation
    • Add the CLAUDE.md clarification (spec_rewrite text) to the COMMIT MESSAGE STRUCTURE section
    • Change the room-attribution call site to prefer Relevant-Rooms / glob-ownership over Originating-Terminal
    • Backtest voice room's 93 prior out-of-lane events: bucket by 'file matches a voice-owned glob' vs 'file matches another room's glob', report mean offPct per bucket, to confirm/falsify this is the dominant driver
    • Add the guard test named in missing_guard in the same commit as the code_rewrite

    What we learned: Voice room's 93 out-of-lane events in 14 days, combined with this commit's own Relevant-Rooms trailer explicitly excluding voice, is evidence that the drift pipeline's room-for-lane attribution is keyed to Originating-Terminal (which terminal the operator physically sat in) rather than to Relevant-Rooms/glob-ownership (which room's files were actually touched) — a divergence CLAUDE.md itself documents as normal and expected, but which the receipt pipeline appears not to account for. This is a new structural finding: prior facts-only interventions for this room never identified a mechanism, only reported outcome drift.

    The missing guard (the deliverable): A test that constructs a synthetic commit whose Originating-Terminal names room X and whose Relevant-Rooms trailer names room Y (X ≠ Y, and the changed file matches only Y's owned-surface glob), then asserts the drift-receipt pipeline's lane/offPct computation is evaluated against room Y, not room X.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not right-sized: litIntent/walksIntent and litReality/walksReality are both exactly 100%, meaning every node the walk visited was counted as lit with no gradation — combined with walkTruncated=true, this reads as the walk being cut short by its depth/time budget before it could produce any faint-vs-solid contrast. Concrete adjustment: either raise --max-depth for this corpus size so the walk completes before truncation, or introduce a sub-100% lit-intensity threshold so 'visited' and 'lit' stop being synonymous. Experiments: Re-run this commit's ballistic walk with truncation disabled/deeper --max-depth and diff offPct against the current 25 — if the saturation is a truncation artifact, offPct should shift measurably. · Backtest the last 93 voice out-of-lane events, partition by 'file matches a voice-owned glob' vs 'file matches another room's glob', and report mean offPct per bucket — directly tests the Originating-Terminal-vs-Relevant-Rooms hypothesis against history. · Add Relevant-Rooms-derived room as a parallel (non-authoritative) label on the next 20 receipt rows and compare its offPct against the current Originating-Terminal-derived offPct to measure which attribution produces tighter, more stable room lanes.

    Measured outcome (deterministic, not self-graded): room mean offPct 21→20 over 4 commits since · dictionary ablation saved 749B

    This event is a false-positive drift signal against voice's lane, not a real quality lapse — if the attribution bug is left unfixed it erodes trust in every future voice-room offPct reading (real drift gets buried in noise), so the premium is restored not by tightening voice's actual output but by re-pointing the receipt's room attribution at Relevant-Rooms so the lane it scores against matches the lane the commit actually claims.

    Open the signed attestation →

59. 2026-09-04 · 🎤 voice · offPct 19 · mistagged

    comms(bookclub): the six-need read rides on every selection, scores and all

    Why it left the lane: The commit message is unusually dense with voice-owned philosophical vocabulary — 'AXIOM 1', 'W6', 'sufficiency contract', 'false-green', and the six canonical need-names CONNECTION/CONTRIBUTION/GROWTH/UNCERTAINTY/CERTAINTY/SIGNIFICANCE — which built strong intent-lane mass toward voice's book/philosophy coordinates. But per PMU DOGFOOD ('reality corpus is ~81% code-as-claim'), the reality corpus was built from the actual changed bytes: `render-email.mjs`'s inline-CSS hex strings (#f7f8fa, #5f6b7a, #1c7a4a, #8a6a10, #a8320d), `PIN_CARRY_FIELDS`, and the test file's `execFileSync`, `mkdtempSync`, `node:test` scaffolding — textbook pipeline/ops identifiers that land in operator/builder coordinates, not voice's books/blog/6needs coordinates. prewalkOverlap=22 against litIntent=119 (~18%) shows the two corpora barely agreed on a seed before the walk even started — a direct fingerprint of prose-register vs code-register mismatch.

    The semantic pull: 'AXIOM 1', 'W6', 'sufficiency contract', 'CONNECTION/CONTRIBUTION/GROWTH/UNCERTAINTY/CERTAINTY/SIGNIFICANCE' (intent side, voice-flavored) vs `PIN_CARRY_FIELDS`, `execFileSync('node', [RENDER, '--passage', pj, '--out', out]`, `border-radius:10px`, and the hex-coded score-bar CSS (reality side, pipeline-flavored) — same commit, two vocabularies, only one of which sits on voice's owned surface.

    Intent–reality gap: The commit claims comms(bookclub) work under Originating-Terminal 🎤 voice, citing voice's own AXIOM/Six-Needs vocabulary at length — but zero bytes of voice's owned surface (books/tesseract/chapters, src/content/blog/*.mdx, src/app/6needs/) were touched. Every changed line is in scripts/bookclub/render-email.mjs and tests/ops/, which is bookclub-pipeline/ops territory the Relevant-Rooms trailer itself names as 🔨 builder and 🎩 operator.

    • Add scripts/bookclub/** and tests/ops/bookclub-*.test.mjs to 🔨 builder's (or 🎩 operator's) owned-surface glob in .workflow/rooms/builder.html
    • Narrow 🎤 voice's vector_keywords in data/rooms.json so 'six needs'/canonical need-names alone don't out-vote file-path evidence when scripts/ or tests/ paths dominate the diff
    • Backfill-scan the last 14 days of voice's 100 out-of-lane events for the same vocabulary-vs-file-surface mismatch signature (AXIOM/Six-Needs prose + scripts/tests-only diff)
    • Add a pre/post-commit check flagging when Relevant-Rooms names a different primary room than Originating-Terminal AND zero Originating-Terminal-owned paths are in the diff
    • Document the 'Six Needs' homonym explicitly in CLAUDE.md next to the pmu-rust/PMU and two-1.2-billions homonyms: taxonomy voice owns (/6needs route) vs literal payload the bookclub pipeline renders

    What we learned: 'Six Needs' is now a homonym exactly like pmu-rust/PMU and the two 1.2-billions: voice owns it as a philosophical taxonomy (/6needs route, AXIOM-heavy commit prose), but scripts/bookclub/render-email.mjs now owns it as a literal per-passage payload it renders — and a commit implementing the payload naturally quotes the taxonomy at length, dragging its room-tag toward voice even when the diff never touches a voice-owned file. This is a NEW class this room's prior 3 sensemaking-unavailable lessons never caught (their offPct trend only worsened, 27→39 and 0→36), because none of them had a diff+message pair to diagnose against.

    The missing guard (the deliverable): A guard that computes, per commit, whether Originating-Terminal's claimed room owns ≥1 changed path (per `.workflow/rooms/<room>.html` globs); if zero owned paths are touched while a DIFFERENT room's globs match every changed path, flag MISTAGGED before the drift receipt escalates — so 'genuinely wrong room' is distinguished from 'genuine cross-lane work' at the source instead of being inferred after the fact.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true with litIntent/litReality both moderate and close in size suggests the walk hit a depth/time budget before full propagation, likely because the commit message's unusual length (AXIOM 1/W6 citations) gives the intent side more to walk than the default budget affords. Adjustment: bump --max-depth one tier for commits whose message length or AXIOM-citation count exceeds a threshold, so intent-side truncation doesn't compound the mismatch. Experiments: Re-seed the reef with the six canonical need-names split into two buckets (voice's taxonomy usage vs bookclub's literal-render usage), re-render this commit's receipt, and check whether offPct drops below the 15 threshold. · Replay the last 5 room-voice commits whose message cites AXIOM 1/W6 at length but whose diff touches only scripts/ or tests/ paths; recompute offPct for each and check whether the pattern is systemic (predict: mean offPct > 15 across all 5). · Track prewalkOverlap specifically for scripts/bookclub/** commits tagged voice vs builder over the next 14 days; if overlap stays low (<25) for voice-tagged ones and higher for builder-tagged ones, that confirms the glob move is the right fix.

    Measured outcome (deterministic, not self-graded): room mean offPct 24→18 over 8 commits since · dictionary ablation saved 1226B

    This is a routing/attribution loss, not a defect recall — the shipped feature and its guard are sound — so the fix is a config correction (owned-surface glob) rather than a code fix, and the room's premium is restored once the glob lands and the next 3-5 bookclub-pipeline commits actually route to builder/operator instead of voice.

    Open the signed attestation →

60. 2026-09-04 · 🔨 builder · offPct 16 · mistagged

    feat(inbox): one name for every inbound channel, and it exits non-zero while any is blind

    Why it left the lane: The whole diff mass (scripts/chat/sync-all.sh, scripts/chat/sync-record.mjs, scripts/chat/sync-verdict.mjs, tests/chat/sync-verdict.test.mjs, .claude/skills/inbox/SKILL.md) sits in paths that match NONE of builder's declared owned-surface globs (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The commit's own Story trailer supplies the pull: 'plumbing over the message spine' — 'plumbing' is the word that drags this toward builder (whose only prior scripts/* claims are build tooling), while the actual content — 'WhatsApp, iMessage, Gmail (sent AND inbox) and Gemini web synced into the single SQLite spine', 'sync-verdict.mjs owns the policy', 'the door to reply to whoever wrote' — is entirely message-channel/inbox-answering work, which is network's (☕ Messages) semantic territory by the room's own persona, not builder's.

    The semantic pull: 'plumbing' + 'sync-all.sh drives every leg' + 'scripts/chat/*' file convention reads as generic infra (builder's normal register: build-book-html.sh, convert-chapters-to-html.mjs) even though the subject matter — WhatsApp/iMessage/Gmail/Gemini inbound channels, 'who wrote in and how to answer them' — is squarely network's declared domain (the Messages persona). Zero of the five changed/new files fall under any of builder's nine owned globs.

    Intent–reality gap: The commit claims 'the work belongs in builder... QC'd toward network, which owns whether a human got an answer' — but 100% of the changed paths (scripts/chat/**, .claude/skills/inbox/**, tests/chat/**) are outside builder's owned surface and inside network's actual subject matter; the receipt measured exactly what the ownership claim denied.

    • Add scripts/chat/**, .claude/skills/inbox/**, tests/chat/** to network's owned-surface glob list in .workflow/rooms/network.html
    • Audit builder's existing scripts/* claims (build-book-html.sh, convert-chapters-to-html.mjs) — confirm nothing else under scripts/ has silently drifted onto builder by the same 'plumbing' pull
    • Add a Relevant-Rooms ordering convention: when 100% of a diff's paths fall in one room's actual glob and 0% in the Originating-Terminal room's glob, the Originating-Terminal MUST be the matching room, not the terminal the operator happened to be in
    • Check whether scripts/pmu/, scripts/ops/, scripts/cog/ have the same orphaned-directory gap (no room's owned-surface glob claims them) before another commit mis-tags itself against one of them

    What we learned: scripts/chat/** (and likely several other scripts/* subtrees) has NO owning room in any .workflow/rooms/*.html glob list — it is an orphan directory, so every commit that touches it gets attributed to whichever terminal the operator was physically in rather than to a fixed semantic owner. The prior three lessons for this room were facts-only (no diagnosis captured) and offPct trended worse across them (16→46, null→26, 15→16); this event suggests the mechanism behind that trend may be the same orphan-directory gap recurring on different unclaimed paths, not three unrelated incidents.

    The missing guard (the deliverable): A commit-time check that computes the overlap between `git diff --name-only` and the Originating-Terminal room's owned-surface globs, and flags/blocks when overlap is 0% while a DIFFERENT room's globs would have matched ≥80% of the same paths — this commit is exactly that case (0% builder match, and scripts/chat/** semantically maps to network) and would have been caught before the receipt needed to fire at all.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent=100% and litReality/walksReality=100% on both sides is unusually saturated — every walked coordinate lit, with no dead-end steps recorded. That can mean genuinely dense connectivity here, or an aperture/max-depth set too shallow to ever hit a non-lighting node, which would mask finer spread differences between intent and reality. Adjustment: re-run with --max-depth increased one ply and check whether the 100%/100% fill ratio holds — if it drops below 100%, the walk was previously too shallow to discriminate; if it holds, the saturation is real and expected for this pairing. Experiments: Re-render 6d051f2b8's receipt after adding scripts/chat/**, .claude/skills/inbox/**, tests/chat/** to network's owned-surface glob and measure whether offPct drops below the 15 threshold — isolates whether the fix is glob-based rather than walk-based. · Add a pre-walk overlap floor check: when prewalkOverlap/litIntent falls below ~10% (as it did here, 8/125), flag the commit for a room-ownership recheck before the full walk runs, rather than only after the receipt lands. · Ablation-test deepening --max-depth by one ply on this commit's reef sha (51e977341413) to see whether the 100%/100% lit/walked saturation is a real ceiling or an aperture artifact.

    Measured outcome (deterministic, not self-graded): room mean offPct 19→18 over 12 commits since · dictionary ablation saved 599B

    A commit whose Originating-Terminal room claims 0% path overlap with its own owned surface is a mistagged-lane loss event, not a code defect — premium on builder's lane is unaffected once scripts/chat/** is reassigned to network, but the orphan-directory pattern behind it (untested across scripts/pmu/, scripts/ops/, scripts/cog/) is an open unpriced exposure until each is explicitly claimed.

    Open the signed attestation →

61. 2026-09-04 · 🔨 builder · offPct 17 · mistagged

    feat(bookclub): the crew call returns to the foot, beside the sticker

    Why it left the lane: The diff's mass is entirely inside render-email.mjs's masthead/footer HTML block: 'sticker-bottom-hiring', the <a href> tracking call, and the inline style scaffolding echo the EXACT strings already used for the top-of-email 'FORK' door and the n5-campaign sticker block a few lines above. gzip-NCD reef seeding pulls those repeated inline-CSS/tracking-URL tokens toward the B (Tactics.Deal/exchange) and C (Operations.Grid/Loop/Flow) coordinates that already own the rest of the email-rendering file's bulk — not toward B3.Tactics.Signal (git-discipline) or C1.Operations.Grid (the commit's claimed lane). The commit message itself is almost entirely prose ABOUT placement strategy (door theory, one-door doctrine, A/B tracking rationale) with only 14 lines of actual diff, so the walk has far more intent-side text than reality-side code to spread across, and the reality lit set lands where the surrounding file's dominant vocabulary already sits.

    The semantic pull: Repeated literal strings pulled the placement: 'iamfim.com' (appears 2x in this diff alone, plus already dozens of times elsewhere in the file), the inline style attribute pattern 'font-size:...;color:...;line-height:...' (identical shape to the sticker div three lines above), and 'track(...)' — the tracking-URL helper call that is the file's single most repeated construct. None of these are git-discipline or commit-pipeline vocabulary; they are bookclub-email-template vocabulary, which is why the encircled reds cluster in B/C Tactics×Operations exchange-and-flow cells rather than in the commit's claimed B3×C1 signal/grid cell.

    Intent–reality gap: The commit message frames this as a room/lane-scoped 'git-discipline' style change (one iteration, one commit, standard bookclub template edit) but the receipt measured the reality lit set spreading across 31 out-of-lane cells at offPct=17 — the work is legitimate, correctly-scoped bookclub-template content, but the RECEIPT's lane assignment (B3.Tactics.Signal × C1.Operations.Grid) never matched what a 14-line HTML-template insertion actually is: an Operations.Flow/Tactics.Deal exchange-copy change.

    • Add an explicit room/lane override in the commit trailer convention for bookclub-email edits — e.g. `Lane-Override: C3.Operations.Flow` when a commit touches scripts/bookclub/render-email.mjs, since that file's dominant vocabulary is exchange/flow copy, not git-discipline.
    • Extend .workflow/rooms/*.html 🔨 builder owned-surface glob to either explicitly EXCLUDE scripts/bookclub/render-email.mjs (routing it to whichever room owns bookclub content) or accept that this file's commits will structurally read as multi-lane and stop scoring them against a single-lane threshold.
    • Audit the 63 out-of-lane events in this room over 14 days for the same file — if render-email.mjs dominates that count, the room-to-file mapping is the systemic cause, not this commit.

    What we learned: A file whose own repeated vocabulary (tracking URLs, inline-style templates, door-copy) is denser and more repetitive than the commit's actual delta will pull the walk toward that file's dominant coordinate regardless of how well-scoped the commit message is — offPct here is measuring FILE IDENTITY, not commit quality, and scripts/bookclub/render-email.mjs is the room's largest source of this exact signature.

    The missing guard (the deliverable): A guard that flags when a commit's reality lit-set breadth (litReality=130, walksReality spanning 31 off-lane cells) is disproportionate to its diff size (14 lines) — a small, single-file HTML-template diff generating a 31-cell spread is itself the anomaly signature worth asserting against, independent of any specific lane.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Too shallow relative to the reef's pull: walkTruncated=true at 2 plies means the walk stopped before it could counterbalance the reef's early lock onto the file's dominant B/C exchange-flow vocabulary. Concrete adjustment: raise max-depth for commits under a size threshold (e.g. <20 line diffs) so the walk gets more plies to potentially resolve the correct single-file lane before truncation, since small diffs are exactly where reef-vocabulary-dominance is most likely to overwhelm actual commit content. Experiments: Recompute this same commit's receipt with prewalkOverlap logged per-candidate-lane (not just a single scalar) to see whether B3×C1 ever ranked above the B/C exchange cells before the walk started, or whether the walk was doomed from seed time. · Run an ablation: strip repeated inline-style-attribute strings from the reef corpus for render-email.mjs specifically and re-render this commit's panel — if offPct drops materially, the template boilerplate (not the commit content) is confirmed as the dominant drag. · Compare offPct across the last 10 commits touching render-email.mjs vs the last 10 touching a lower-repetition file of similar diff size in the same room, to quantify whether file-vocabulary-density predicts offPct independent of commit quality.

    Measured outcome (deterministic, not self-graded): room mean offPct 20→38 over 4 commits since · dictionary ablation saved 1009B

    A mistagged-lane event on a well-scoped, low-risk template edit shouldn't move this lane's loss ratio on its own, but 63 out-of-lane events in 14 days for one room signals the room-to-file mapping itself is underwritten wrong — restoring premium here means fixing the lane assignment upstream, not re-scoring individual commits like this one.

    Open the signed attestation →

62. 2026-09-04 · 🔨 builder · offPct 21 · structural

    fix(gemini): ask once, not 402 times — and a blind corpus exits non-zero

    Why it left the lane: No encircled-region narrative is cached for this sha, so the only placement evidence available is the diff's own paths and vocabulary. All 137 changed lines land in scripts/chat/gemini-web-all.sh, scripts/ops/initiative-checkpoint.mjs, scripts/ops/initiative-loop.sh and tests/chat/gemini-preflight.test.mjs — zero lines touch src/app/, src/components/, src/lib/, public/, or any other glob on 🔨 builder's declared owned surface. The pull terms are pure operations/pipeline nouns — 'PREFLIGHT', 'corpus', 'checkpoint', 'GEM_AGE_D', 'cp_ corpus --ok'/'--fail', the new `corpus:` stage in STAGES — which gravitate to loop-staging/checkpoint vocabulary, not to anything builder's owned-surface list claims. Reality landed exactly where the vocabulary says it must; the mismatch is that builder's lane structurally excludes that whole space.

    The semantic pull: 'GEM_AGE_D', 'cp_ corpus', 'STAGES.corpus', 'chrome-cdp.mjs --status', 'initiative-loop.sh' are all Operations-axis terms (staging/checkpointing a pipeline run) with no analog anywhere in builder's UI/component/public glob — there is nothing in the diff pulling toward src/app or src/components, so the whole commit's mass sits off builder's declared lane by construction, not by drift within a shared region.

    Intent–reality gap: The commit message and trailers claim this work as 🔨 builder's (Originating-Terminal: builder, Relevant-Rooms including builder), but the receipt measured 100% of the touched surface sitting in scripts/chat/, scripts/ops/, and tests/chat/ — paths that appear nowhere in builder's owned-surface list. The commit is a correct, well-scoped infra fix; the lane it was filed under is the wrong one.

    • Add scripts/ops/**, scripts/chat/**, tests/chat/**, tests/ops/**, docs/ops/mailbox/** to a room's owned-surface glob in .workflow/rooms/*.html (naming fit: 🎩 operator, given 'Kitty Operator' already carries the automation/ops persona) so Relevant-Rooms scoring stops defaulting to Originating-Terminal for this content class.
    • Run the /recontext skill against the current tree to enumerate every other UNOWNED or CONTESTED path glob before it produces more chronic false-positive drift events like this one.
    • Add a coverage guard (e.g. tests/ops/owned-surface-coverage.test.mjs) asserting every top-level scripts/** and tests/** directory matches at least one room's owned-surface declaration.
    • Wire the drift-receipt lane computation to weight offPct against the full Relevant-Rooms set (not solely Originating-Terminal) so a commit correctly filed under an unclaimed surface doesn't inherit whichever terminal happened to be open.
    • Audit the room's other 56 out-of-lane events from the last 14 days for the same scripts/ops·scripts/chat·tests/** signature — if most share it, builder's high offPct trend is mostly a glob-coverage artifact, not genuine drift.

    What we learned: The 9-room owned-surface taxonomy has a real gap: infra/pipeline paths (scripts/ops/**, scripts/chat/**, tests/**) are claimed by NO room's glob, so any commit touching them — regardless of which terminal made it — cannot land inside a claimed lane at all. Every such commit will read as out-of-lane by construction until the glob coverage is extended, which means this room's offPct trend is partly measuring a taxonomy hole rather than builder-specific drift.

    The missing guard (the deliverable): A coverage test asserting every directory under scripts/** and tests/** matches at least one room's owned-surface glob in .workflow/rooms/*.html — an orphan path should fail the test loudly rather than silently defaulting to Originating-Terminal at commit time.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent==walksIntent (139==139) and litReality≈walksReality (142 vs 142) — every walk step lit a distinct new cell with no revisits, consistent with a shallow, single-ply walk per seed rather than deep recursive clustering. For a diff this size (137 lines, 4 files) that's plausibly right-sized in breadth, but concretely worth testing: re-run with --max-depth+1 on this commit's seeds to see whether ops-vocabulary anchors (checkpoint/corpus/preflight) connect through shared rows/columns to anything in builder's declared lane — if depth was truncated at ply 1, some legitimate connectivity may be going undiscovered. Experiments: Re-seed prewalkOverlap using the full Relevant-Rooms trailer (builder+navigator+operator) instead of only Originating-Terminal, re-render this commit's panel, and measure whether offPct changes. · Add scripts/ops/**, scripts/chat/**, tests/chat/** to a room's owned-surface glob, re-render sha 727e802dd, and measure whether offPct drops below the 15 threshold. · Re-run the ballistic walk on this commit's seeds with --max-depth+1 and compare litIntent/litReality counts to test whether the near-1:1 lit:walk ratio was a depth-truncation artifact rather than a true right-sizing.

    Measured outcome (deterministic, not self-graded): room mean offPct 19→16 over 5 commits since · dictionary ablation saved 982B

    This event is a lane-ownership gap, not a defect — the fix is well-scoped and well-guarded, so premium risk on the underlying work is low, but builder's loss-ratio reads artificially high until the owned-surface globs are extended to cover scripts/ops/, scripts/chat/, and tests/, which is what actually restores it.

    Open the signed attestation →

63. 2026-09-04 · 🎩 operator · offPct 37 · mistagged

    fix(chat): the WhatsApp copy is bounded — a modal nobody answers was an infinite hold, and it disarmed the gate

    Why it left the lane: The diff's vocabulary is entirely chat/security domain — "TCC", "MODAL", "tcc-gate.mjs --deny whatsapp-store", "whatsapp-store.granted", "mkfifo", "backoff" — and none of it overlaps the operator room's declared surface terms (trust-debt, pmu/pipeline, dashboard, accounting, ORTHOGONAL_HEATMAP). With zero shared vocabulary between the diff and operator's owned-surface globs, gzip-NCD reef seeding has nothing local to anchor to and the walk lands its mass on whichever reef seeds DO share vocabulary — which, per the commit's own Relevant-Rooms trailer, are ☕ network (chat pipeline) and 🔒 vault (TCC/consent gating). The commit correctly named those rooms itself; the receipt is just scored against the wrong one.

    Intent–reality gap: The commit message claims a correct, well-scoped fix to the WhatsApp/TCC copy pipeline — and it is one, with a 9/9 guard. The gap isn't in the work, it's in the lane: this sha is scored against 🎩 operator's owned surface (trust-debt/pmu-pipeline/dashboards/accounting), but scripts/chat/ and tests/chat/ never appear in that surface at all, so the receipt reads a correctly-placed commit as 37% off-lane.

    • Add scripts/chat/**, tests/chat/** to ☕ network's owned-surface glob in .workflow/rooms/network.html — currently unclaimed by any room, which is why the reef has no home lane for chat-pipeline commits
    • Add scripts/chat/tcc-gate.mjs and TCC/consent-boundary paths to 🔒 vault's owned-surface glob in .workflow/rooms/vault.html
    • Patch the CLAUDE.md COMMIT MESSAGE STRUCTURE "Why" note per spec_rewrite so the accepted terminal/file divergence stops reading as drift
    • Ship the lane-attribution guard named in missing_guard
    • Audit 🎩 operator's other 14 recent out-of-lane events for the same terminal-of-convenience pattern before assuming they're genuine content drift

    What we learned: CLAUDE.md itself already sanctions Originating-Terminal diverging from file-relevance ("that's fine and worth recording") — but the drift-receipt pipeline has no mechanism that reads that sanction; it scores every commit against whatever room the physical terminal happened to be, so a room used as a general-purpose staging terminal (as 🎩 operator apparently is, at 15 events/14 days) will structurally out-score its own declared lane regardless of code quality, because the receipt and the commit-trailer convention were never wired to agree.

    The missing guard (the deliverable): A test asserting the receipt pipeline's lane baseline prefers the Relevant-Rooms glob-match room over the Originating-Terminal room whenever the match confidence is decisively higher — SMOKE: a fixture commit whose Originating-Terminal is room A but whose changed files glob-match room B's owned surface at high confidence must score offPct against B, not A; NO SMOKE: a commit whose Originating-Terminal room's own surface is the best glob match must still score against that room unchanged.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent/walksIntent and litReality/walksReality both sit at ~100% (130/130, 134/134) — every walked node lit, meaning the walk is fully saturated rather than exploring-and-pruning. That saturation plus the near-equal walk counts on both lanes suggests the walk is hitting a time/ply budget wall symmetrically rather than converging naturally. Adjustment: rerun with --max-depth increased for this commit's reef and check whether litReality keeps climbing past 134 — if it plateaus, the walk is right-sized; if it keeps growing, the current depth is truncating real signal on both sides equally. Experiments: Re-seed this commit's reef using Relevant-Rooms-weighted glob matching instead of Originating-Terminal-only, re-render the receipt, and measure whether offPct drops toward network/vault's own baseline offPct · Ablate the reef vocabulary for TCC/modal/grant/backoff terms (present/absent) against operator's declared surface and measure the offPct delta directly attributable to that vocabulary · Log Originating-Terminal-vs-glob-match-room agreement across the room's last 15 out-of-lane events and report what fraction is this same mistagged-terminal pattern vs genuine content drift

    Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 6 commits since · dictionary ablation saved 942B

    The work itself is sound and guarded (9/9), so this event should not raise 🎩 operator's true loss-ratio — but until lane attribution is fixed, every terminal-of-convenience commit like this one will keep inflating the room's measured offPct, and a premium computed off that inflated number is pricing a room for drift it didn't actually produce; restoring accuracy requires the Relevant-Rooms-weighted lane fix, not a reduction in commit scope.

    Open the signed attestation →

64. 2026-09-04 · 🎤 voice · offPct 51 · structural

    test(claims): four detectors behind one assertion was measuring the aggregate

    Why it left the lane: The diff's actual vocabulary — `FABRICATIONS`, `fire:`/`corrected:` regex pairs, `assert.deepEqual(blind, [])`, `scanBody('hypothetical-post.mdx', ...)`, `paragraphs(raw)` — is test-harness/mutation-verification language living in `tests/claims/no-boriscap-fabrications.test.mjs`. None of that path or vocabulary intersects 🎤 voice's declared owned surface (`books/tesseract/chapters/*.md`, `src/content/blog/*.mdx`, `docs/05-content/blog/scratchpad/`, `decentralizedtalk/slideData.ts`, `src/app/decentralized*`, `src/app/6needs/`, `scripts/voice-*-check.sh`). The mass landed squarely in structural/verification lattice territory (assertion logic, mutation-tested detectors) rather than prose-authorship territory, which is exactly what the receipt's offPct=51 is reporting.

    The semantic pull: Pull terms: 'FOUR REGEXES, ONE `>= 3` ASSERTION', 'Verified by mutation, one at a time', 'A CORRECTION MARKER', 'the aperture stays the blank-line block' — these are guard-engineering constructs (aperture, mutation, assertion granularity) that the reef's gzip-NCD seed correctly reads as architect/laboratory-shaped prose (the Relevant-Rooms trailer itself names 📐 architect and 🧪 laboratory, never voice), not as blog/book content-authorship prose.

    Intent–reality gap: The commit message claims a room-scoped fix (test guard correctness, mutation coverage, marker-word precision) with zero reference to any blog/book/chapter content — yet it is attributed to 🎤 voice, whose owned surface is entirely content files. The receipt confirms the mismatch directly: 51% of lit reality mass sits outside the claimed lane, and the commit's own Relevant-Rooms trailer already names the correct lane (architect/laboratory) without voice appearing in it.

    • Add `tests/claims/*fabrication*.test.mjs` to voice's owned-surface glob in the room definition (`.workflow/rooms/voice.html` or equivalent).
    • Ship `scripts/voice-boriscap-check.sh` wrapping the mutation-tested guard, matching the existing `scripts/voice-*-check.sh` pattern voice already owns.
    • Cross-link the guard from CLAUDE.md's fabrication/claims section so architect/laboratory (named in Relevant-Rooms) and voice (named in Originating-Terminal) both know who's accountable for it.
    • Audit other rooms for the same shape: a guard/test file that protects content X but lives in tests/ or scripts/ outside X's owning room's glob.
    • Extend the individually-asserted-detector pattern this commit introduced to any other multi-regex guard in tests/claims/ or tests/voice/ that still hides behind one aggregate `>= N` assertion.

    What we learned: A content-safety guard (fabrication/claim detection) is conceptually owned by the room whose content it protects, but physically lives in `tests/`, a directory no content room's owned-surface glob has ever named — so any terminal that does legitimate guard-hardening work for its own domain will read as out-of-lane by construction until the glob is extended to follow the guard, not just the prose.

    The missing guard (the deliverable): A room-lane consistency check: for every commit, assert that the touched paths intersect at least one glob in the Originating-Terminal room's owned surface OR that the Relevant-Rooms trailer already lists a room whose glob does match — and when neither holds, fail with 'ROOM CONFIG GAP: <room> touched <path>, which matches no glob in any listed room' rather than silently letting the drift receipt carry the whole signal.

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Both lanes are 100% lit-to-walked (no dead branches at all) and walkTruncated=true — that combination means the walk hit its depth/budget ceiling before naturally converging, not that it saturated healthily. Concrete adjustment: re-run this commit's walk with `--max-depth` raised past the current ceiling and check whether litReality/litIntent counts grow past 126/118 and offPct drops — if it doesn't move, truncation wasn't the driver and the seed mismatch (prewalkOverlap=7) is confirmed as the whole story. Experiments: Re-seed the reef restricted to only voice's owned-surface files vs. the actual touched files in this commit, and measure whether prewalkOverlap stays near 7 — if so it confirms structural mismatch rather than a fixable seed problem. · Replay sha 8aa530e75 with `--max-depth` doubled and diff the resulting offPct against 51 to isolate how much of this event was truncation vs. seed disjunction. · Ingest the commit's own Relevant-Rooms trailer as an additional reef seed lane (architect+laboratory+operator globs) and measure offPct against THAT lane — if it drops sharply, the trailer is already carrying the correct-lane signal the pipeline isn't using.

    Measured outcome (deterministic, not self-graded): room mean offPct 21→20 over 13 commits since · dictionary ablation saved 1171B

    A structural glob gap masquerading as content drift inflates this lane's measured loss ratio with false positives, which erodes trust in the receipt itself; extending voice's owned-surface glob to its own guard files (plus the wrapper script) restores the premium by making the lane assignment match the actual, legitimate work being insured.

    Open the signed attestation →

65. 2026-09-04 · 🎤 voice · offPct 37 · mistagged

    feat(milestones): the board learns two things the dinner taught it

    Why it left the lane: Every changed path in this commit is Operations/Argument-track ledger vocabulary — data/milestones-market-contact.json, data/milestones-next.json, data/milestones-progress.json, data/milestones.json, tests/ops/milestones-shortlex.test.mjs, and inside the diff itself the assertions on b.placed/b.saidWell/b.witnessed and the comment language 'evidenced against ledgers rather than files we author' — none of it is book/blog prose. 🎤 voice's owned surface (books/tesseract/chapters/*.md, src/content/blog/*.mdx, decentralizedtalk/slideData.ts, src/app/6needs/) has zero glob overlap with any changed file. The receipt's litReality=122 spread lands entirely in the Operations/Argument coordinates the milestones board occupies, not in voice's book/blog coordinates, which is exactly the offPct=37 the detector reports.

    Intent–reality gap: The commit message is accurate about what it did (milestones board updates + a stale-pin correction) and even self-declares the correct owners in its own trailer — Relevant-Rooms: 🎩 operator, 💰 network, 📐 architect — but Originating-Terminal reads 🎤 voice, and the receipt pipeline appears to lane-check against Originating-Terminal's owned surface rather than the Relevant-Rooms it named itself.

    • Update CLAUDE.md's COMMIT MESSAGE STRUCTURE section with the Relevant-Rooms-wins-for-lane-checking rule
    • Patch the receipt's room-lane resolver to union Relevant-Rooms' owned globs when present
    • Confirm `.workflow/rooms/operator.html`, `network.html`, `architect.html` actually declare `data/milestones-*.json` and `tests/ops/**` in their owned-surface globs — if they don't, add them there first, since the resolver fix is inert without real glob coverage
    • Backfill-check the 90 prior 🎤 voice out-of-lane events over the last 14 days against this hypothesis: how many carry a Relevant-Rooms trailer that excludes voice, vs how many are genuine book/blog-content drift
    • Add a firing test per EVERY GUARD SHIPS WITH A FIRING TEST: feed the resolver this commit's trailers and assert offPct drops when Relevant-Rooms is honored, and assert it stays at 37 on a synthetic commit with no Relevant-Rooms trailer at all

    What we learned: The receipt pipeline's lane check binds to Originating-Terminal, but CLAUDE.md's own commit-trailer convention explicitly designs Originating-Terminal and Relevant-Rooms to diverge — meaning the detector is measuring a dimension (physical terminal) the convention never promised would predict content, while ignoring the dimension (Relevant-Rooms) the commit author computed specifically to name where the work belongs. The 90-events-in-14-days figure for this room is the tell: this isn't voice room content-drifting, it's voice room being used as a physical workspace for ops/argument bookkeeping while correctly self-reporting elsewhere in its own trailer, and the pipeline never reads that self-report.

    The missing guard (the deliverable): A test asserting the receipt's lane-membership computation prefers Relevant-Rooms' owned-glob union over Originating-Terminal's owned glob whenever a Relevant-Rooms trailer is present — fed a real commit (like this one) whose changed files match Relevant-Rooms but not Originating-Terminal, and asserting offPct falls under threshold once the union is used.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not right-sized — a 1:1 walked:lit ratio on both sides means the walk floods every reachable node rather than pruning any of it, so it adds no discriminating power over the raw seed placement. Concrete adjustment: lower `--max-depth` by one ply for commits where walksX==litX on both sides for three consecutive commits (a saturation signature), and check whether offPct separates real content-drift from mistagged-terminal cases once the walk stops just mirroring the seed. Experiments: Recompute prewalkOverlap across the last 20 out-of-lane 🎤 voice commits and correlate it against whether Relevant-Rooms excludes Originating-Terminal — if the correlation is strong, that's the mechanical confirmation this diagnosis needs before the resolver patch ships · Replay this commit's reef seed step twice: once keyed to Originating-Terminal's glob set, once to the Relevant-Rooms union, and diff litReality/offPct between the two runs to get a real before/after number instead of a hypothesis · Add a saturation counter to the walk output (walksX vs litX delta) and flag any commit where the delta is zero on both sides three commits running, as a signal the walk step is currently inert for that room

    Measured outcome (deterministic, not self-graded): room mean offPct 22→21 over 8 commits since · dictionary ablation saved 737B

    A chronic mistagging pattern (90 out-of-lane events in 14 days for one room) reads to an underwriter as an uncalibrated sensor, not as 90 real drift incidents — the premium hit is on the RECEIPT's credibility, not on this commit's content, and it's restored by making the lane check honor Relevant-Rooms rather than by disputing this commit's placement.

    Open the signed attestation →

66. 2026-09-04 · 📐 architect · offPct 17

    feat(comms): candidate WhatsApp replies, drafted against the measured room rules

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 22→7 over 6 commits since

    Open the signed attestation →

67. 2026-09-04 · 🔨 builder · offPct 28 · mistagged

    feat(initiative): phase 3 — twelve is a review unit, and the 45-minute close fires the loop

    Why it left the lane: Every changed file in this commit — scripts/ops/initiative-dispatch.mjs, scripts/ops/initiative-loop.sh, scripts/ops/session-close-cycle.sh, tests/ops/initiative-dispatch.test.mjs — sits under scripts/ops/, which appears in NO glob on builder's owned-surface list (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, scripts/voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The vocabulary that pulls the reality walk away from builder's lane is exactly the ops-pipeline terms saturating the diff: 'phase 1 · mine', 'phase 2 · frame', 'phase 3 · dispatch', 'quiet gate', 'docs/comms/self/drafts', 'framed-*.json', '( nohup … & ); disown' — none of it touches src/app or public. The commit's own Relevant-Rooms trailer already names 🎩 operator, not builder, as the file-relevant room; the offPct=28 is what happens when the lane check is scored against Originating-Terminal (which terminal was physically open) instead of against that trailer.

    The semantic pull: Strings 'initiative-dispatch.mjs', 'initiative-loop.sh', 'session-close-cycle.sh' under scripts/ops/, plus 'quiet-gate.mjs --check', 'framed-', 'outbox', 'dead-letter', 'docs/comms/self/drafts' — an operations/mailbox vocabulary cluster with zero overlap against builder's UI/build-tooling glob set, dragging the reality walk into operator-owned territory while intent stayed pinned to builder by the Originating-Terminal trailer alone.

    Intent–reality gap: The commit is tagged Originating-Terminal: 🔨 builder and scored against builder's lane, but 100% of the diff (326 insertions across 4 files) lands in scripts/ops/ + tests/ops/ — a surface builder does not own and operator does; the commit's own Relevant-Rooms trailer already says so, but the offPct math didn't read it.

    • Confirm the room-glob file for 🎩 operator (`.workflow/rooms/operator.html`) already lists `scripts/ops/**`; if not, add it in the same commit as any lane-resolver fix.
    • Confirm builder's owned-surface file does NOT list `scripts/ops/**` (per the surface given above, it correctly doesn't — document this as intentional, not an oversight).
    • Locate the script that computes per-commit offPct/lane placement and patch it to resolve the scored room via Relevant-Rooms[0] (or the best owned-glob match) rather than the Originating-Terminal trailer.
    • Add a regression test: a synthetic commit with Originating-Terminal=builder and 100% of changed paths under scripts/ops/** must score against operator's lane, not builder's.
    • Re-check builder's room offPct trend after the fix — the prior three interventions moved it null→26, 16→46, 15→16, and this fix should show that oscillation flattening once ops-pipeline commits stop being charged to builder.
    • Add a line to CLAUDE.md's Originating-Terminal/Relevant-Rooms description naming the lane-resolver script by path so the rule is checkable, not just prose.

    What we learned: The offPct/lane-scoring pipeline is keyed off Originating-Terminal (which terminal was physically open), not Relevant-Rooms (whose files were actually touched) — and CLAUDE.md explicitly sanctions the two diverging, which means every commit made from the builder terminal for genuinely operator-owned work (scripts/ops/*) structurally inflates builder's drift trend regardless of code quality. This room's rising/oscillating offPct across the last three interventions (16→46, 15→16, now 28) is plausibly a symptom of this same mistagging recurring, not of builder's actual work drifting.

    The missing guard (the deliverable): A test asserting the lane-resolver scores a commit against Relevant-Rooms[0]/best-owned-glob-match rather than Originating-Terminal — fixture: a synthetic commit with Originating-Terminal=builder and changed paths 100% under scripts/ops/**, asserting the scored room is operator (or whichever room's glob matches), not builder.

    Detector probe: signal attribution — gzip reef 20% · ballistic walk 80%. litIntent/walksIntent = 123/123 and litReality/walksReality = 142/142 — a 1.0 hit-rate on both sides means the walk never revisited a cell, which reads as saturated rather than under-used, but walkTruncated=true says it was cut off by the depth/step budget before intent could catch up to reality's larger footprint. Concrete adjustment: raise --max-depth for commits over ~300 insertions (this one, 326) so intent-side lit count has room to converge toward reality-side lit count before the budget stops the walk, rather than truncating mid-divergence and reading the gap as offPct. Experiments: Re-run this commit's reef seeding with Relevant-Rooms (operator) terms included alongside Originating-Terminal (builder) terms, and measure whether prewalkOverlap rises above 3 — if it does, seed composition is part of the fix, not just the lane resolver. · Raise --max-depth for diffs >300 insertions and replay this commit; measure whether litIntent (123) converges closer to litReality (142), which would show truncation, not genuine drift, explains part of offPct=28. · Add scripts/ops/*-vocabulary ('initiative', 'dispatch', 'mailbox', 'quiet gate', 'framed') to the operator-room reef corpus and replay this commit's receipt; measure whether offPct drops independent of the room-glob config fix, isolating how much of the 28 was reef-corpus-thin versus lane-mis-scored.

    Measured outcome (deterministic, not self-graded): room mean offPct 17→14 over 10 commits since · dictionary ablation saved 1278B

    Scored as-is against builder's lane this event inflates builder's loss-ratio for work that is entirely correct and entirely operator's, so premium is restored not by touching the shipped code but by re-keying the lane check to Relevant-Rooms/owned-glob match — a zero-cost config + resolver fix, not a code-quality fix.

    Open the signed attestation →

68. 2026-09-04 · 🔨 builder · offPct 41 · mistagged

    test(chat): the redirect guard carries its own seen-to-fail receipt

    Why it left the lane: The diff's mass is a git-history forensics + fixture-provenance move: execFileSync('git',['show',...]), 'historical', 'the fixture is carried inline, drawn through git show from the commit that shipped the defect', 'wrong fixture', 'could not read the historical defect'. That vocabulary is Operations/verification-of-past-state language (git archaeology, provenance chain, refuse-vs-fail semantics), not Signal (B3) content — it walks straight into the Grid/Loop/Flow × Speed/Deal cells the receipt circled (topology, velocity, throughput, exchange), because 'pull historical bytes and re-run the same predicate against two states' reads to the reef as an operations/exchange pattern, not a communications-guard pattern.

    Intent–reality gap: The commit message frames this entirely as a chat/gmail-redirect TEST fix ('the redirect guard carries its own seen-to-fail receipt') and the room claim is 🔨 builder/B3.Tactics.Signal, but the actual work is a generic git-provenance harness pattern (extract predicate once, apply to live file + git-show'd historical bytes, fail loudly on git error) that has nothing semantically to do with gmail/oauth/redirect and everything to do with operations-grid verification machinery.

    • Add a one-line scope note atop the new SMOKE test explaining it is a LOCAL instance of the repo-wide bootstrap-fixture pattern (points to guard-seen-to-fail.mjs's own bootstrap-case clause)
    • Extract the execFileSync/git-show/assert.fail block into tests/lib/historical-fixture.mjs and reuse it here
    • Cross-link this test's header comment to tests/ops/guard-seen-to-fail.test.mjs so the pattern has one canonical home
    • Update .workflow/rooms/*.html owned-surface globs so tests/lib/ (once created) is claimed by whichever room owns testing-infrastructure, not implicitly builder
    • Add a CLAUDE.md pointer under the guard-seen-to-fail HARD RULE noting the bootstrap-case discharge now has a second, reusable implementation

    What we learned: A test file whose commit message names ONE narrow subject (a redirect guard) can still smuggle in a REUSABLE, general-purpose testing-infrastructure pattern (the git-show fixture harness) that the reef correctly recognizes as Operations/Grid content — the drift here isn't bad work, it's that a genuinely repo-wide pattern was born inline in a single-purpose file instead of being named and extracted as infrastructure at birth, so its own commit message couldn't honestly claim just one lane.

    The missing guard (the deliverable): A structural check (e.g. tests/ops/no-inline-general-fixture-harness.test.mjs) that flags when a test file under tests/<domain>/ contains a self-contained 'try { execFileSync(git show...) } catch { assert.fail }' block NOT imported from tests/lib/ — forcing any repo-wide git-provenance-fixture pattern to be extracted before it can proliferate inline across unrelated domain test files.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. walkTruncated=true with litReality(134) > litIntent(110) and only 2 plies suggests the walk hit its budget before fully separating the gmail-signal terms from the git-archaeology terms — it's running shallow-and-saturated rather than shallow-and-precise. Concrete adjustment: bump --max-depth by one ply specifically when walkTruncated=true and litReality/litIntent ratio exceeds ~1.2, so the walk gets one more hop to disambiguate before truncation locks in the spread. Experiments: Re-run this exact commit's ingest with --max-depth+1 and compare offPct — if it drops meaningfully, truncation (not real drift) explains part of the 41% · Ablate the phrase 'git show' and 'execFileSync' from the reality corpus for this commit only and re-measure offPct, to isolate how much of the 28 red regions trace to the provenance-harness vocabulary specifically vs the gmail-redirect predicate code · Compare prewalkOverlap across the last 10 builder-room commits with walkTruncated=true vs false to check whether truncation itself correlates with high offPct room-wide (would explain the room's 56-events-in-14-days pattern)

    Measured outcome (deterministic, not self-graded): room mean offPct 19→14 over 3 commits since · dictionary ablation saved 1224B

    A guard whose own verification tooling (guard-seen-to-fail.mjs) structurally cannot run against it is a documented coverage gap in this lane's underwriting — the bootstrap discharge here restores insurability for THIS guard, but the missing_guard is what prevents the same gap from reopening on the next inline-fixture-harness commit.

    Open the signed attestation →

69. 2026-09-04 · 🔨 builder · offPct 19 · mistagged

    feat(gift): you can't buy your own alibi — send a key to somebody else

    Why it left the lane: The commit's prose is written entirely as a product-feature narrative (gift page, API route, floater component — all inside 🔨 builder's owned src/app/ and src/components/ surface), so the intent corpus the reef ingested carries almost no semantic mass for the one hunk that actually pulls the reality walk sideways: the ~20-line insert into scripts/bookclub/render-email.mjs (the 'Send a key' block, the track('https://thetadriven.com/gift', day, 'gift-key') call, the '#faf7f0/#e0b040' comms styling). That file sits in the newsletter/outreach pipeline, not in any glob builder owns, and it is real, deliberate work (the Story trailer even says the operator asked for the CTA) — it just was never named in the body prose that seeds the intent lane, only in the Relevant-Rooms trailer (☕ network), which the reef does not read as intent text. tests/components/gift-license-floater.test.mjs (tests/ is also outside builder's listed globs) is a smaller second contributor of the same shape.

    The semantic pull: Exact strings dragging reality off the claimed lane: 'GIFT A KEY' / 'Send a key' / track('https://thetadriven.com/gift', day, 'gift-key') / the inline #faf7f0;border-left:3px solid #e0b040 block — all inside scripts/bookclub/render-email.mjs, none inside src/app/ or src/components/. No region-coordinate narrative was cached for this sha, so this is grounded in file-path/glob divergence rather than invented lattice coordinates.

    Intent–reality gap: The message claims a single-lane builder feature ('the request route', 'the mint already supports this', 'the checkout button') and never once names scripts/bookclub/render-email.mjs; the receipt measured 19% of lit reality mass landing in exactly that unmentioned file — the intent text and the diff disagree about how many lanes this commit touches, even though the commit's own Relevant-Rooms trailer (🎤 voice, 🔨 builder, ☕ network) already knew it was three rooms wide.

    • Extract the 'Send a key' HTML fragment out of scripts/bookclub/render-email.mjs into its own module (see file_movements) so the builder-committed call site is a one-line import, not the comms markup itself
    • Add scripts/bookclub/render-email.mjs (or the new blocks/ dir) to ☕ network's owned-surface glob in the room config so future edits there compute against network's lane, not whichever room happens to commit them
    • Update CLAUDE.md's COMMIT MESSAGE STRUCTURE section to note: when Relevant-Rooms names a room other than Originating-Terminal, the body prose must name the specific file(s) that justify it
    • Add tests/ (or tests/components/) to a room's owned-surface list — it currently belongs to none, so any commit touching it accrues unexplained offPct
    • Add a short comment atop the extracted gift-key block module naming which commit/thesis it implements, so a future reef ingest has semantic mass at the source rather than only at the call site

    What we learned: The Relevant-Rooms trailer is a documented, correct multi-room signal (CLAUDE.md already says divergence between Originating-Terminal and file-relevance is 'fine and worth recording'), but the reef's intent-lane computation reads commit BODY prose for semantic mass and never consumes the trailer — so a commit that correctly disclosed its own multi-room scope in metadata still registers as fully out-of-lane, because the disclosure lived in the wrong half of the message.

    The missing guard (the deliverable): A pre-commit advisory (warn-only) that, when a commit's diff touches files outside the Originating-Terminal room's owned-surface globs but inside another room named in Relevant-Rooms, checks the commit body prose for at least one sentence naming that file or its purpose — and prints a reminder to add one if absent.

    Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. Right-sized, not shallow: both sides are fully lit (litIntent=walksIntent, litReality=walksReality) with no truncation, so the walk ran to completion rather than cutting off early. The lever here is on the INPUT side, not depth — grow litIntent's coverage by feeding the intent corpus the disclosure sentence (spec_rewrite) so the walk has real intent-side mass over the render-email.mjs coordinates instead of having to reach them purely by reality-side propagation. Experiments: Re-render this exact sha's receipt against a scratch commit message that adds the disclosure sentence (without landing it) and check whether offPct drops below the 15 threshold — isolates whether the fix is in the prose alone or needs the glob change too · Pull prewalkOverlap for the last 20 🔨 builder out-of-lane events and check whether it is consistently low (<10) specifically when the diff touches an ungloobed file like scripts/bookclub/; if so, low prewalkOverlap becomes a cheap early-warning signal before the full walk runs · Re-walk this same sha after adding scripts/bookclub/** to ☕ network's owned-surface glob and measure whether offPct's denominator (what counts as 'in lane' for this commit) changes independent of any prose edit — tests whether part of the drift is an artifact of an unowned file rather than true semantic divergence

    Measured outcome (deterministic, not self-graded): room mean offPct 20→20 over 8 commits since · dictionary ablation saved 1159B

    The underlying work is legitimate and operator-directed rather than defective, so this event should nudge 🔨 builder's loss-ratio only mildly — but it keeps feeding a room already running 61 out-of-lane events in 14 days with a worsening trend, and premium is restored by moving scripts/bookclub/render-email.mjs under ☕ network's glob and disclosing cross-lane edits in body prose rather than only in the trailer.

    Open the signed attestation →

70. 2026-09-04 · 🔒 vault · offPct 24 · mistagged

    fix(gem): gemstart never started — an empty array is unbound on the bash macOS ships

    Why it left the lane: The diff itself (scripts/ops/gem-loop.sh's FRAME_ARGS/unbound-variable fix, tests/ops/gem-loop.test.mjs's scratchTree/spawnSync firing tests) shares zero vocabulary with vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/) — there is no gzip-NCD content overlap to chase here. The pull is a METADATA pull, not a semantic one: the commit's own trailer literally reads 'Originating-Terminal: 🔒 vault' (the terminal the operator was physically sitting in), and the receipt pipeline appears to key its room-for-offPct attribution off that string rather than off 'Relevant-Rooms: 🔨 builder, 🎩 operator, 🧪 laboratory' — the field this same commit correctly populated with the rooms whose owned-surface globs the diff actually matches. litIntent/litReality (123/125) and walksIntent/walksReality (123/125) track each other almost exactly, which is the tell: the walk explored and lit nearly identical volume on both sides, so the 24% off-lane share isn't a walk-behavior artifact, it's the wrong room's glob being used as the intent lane in the first place.

    The semantic pull: The literal trailer token 'Originating-Terminal: 🔒 vault' is what pulled this event into vault's lane-scoring, not any string in the actual diff. No diff term (FRAME_ARGS, unbound variable, initiative-loop.sh, quiet-gate) has any affinity to vault's patent/legal/data-room vocabulary.

    Intent–reality gap: The commit's own trailers already state the truth — Relevant-Rooms names builder/operator/laboratory and does not include vault at all — but the receipt still scored this event's offPct against vault's owned-surface glob (which the diff touches zero files of), producing an off-lane reading for a commit that was never claiming vault's lane to begin with.

    • grep the receipt pipeline for every site that reads the 'Originating-Terminal' trailer to resolve a room, not just the one e94a8cd14 patched — there are at least two independent consumers of this trailer and only one appears fixed
    • add a firing test using bb6c10607's own trailers as fixture: SMOKE asserts today's pipeline scores offPct against vault despite zero vault-glob matches; NO SMOKE asserts the patched pipeline scores it against builder/operator/laboratory instead
    • extend the intervention-brief generator to print BOTH Originating-Terminal's and Relevant-Rooms' owned-surface globs side-by-side so a reader sees the divergence instead of only ever seeing one room's globs
    • re-walk the room's other 24 out-of-lane vault events from the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms divergence — if most share it, this single pipeline fix retroactively explains most of fdfcf4172's 11→24 offPct climb
    • tighten CLAUDE.md's COMMIT MESSAGE STRUCTURE section to state explicitly that automated lane/offPct scoring MUST read Relevant-Rooms, never Originating-Terminal — the current text implies but never states this, which is exactly the gap that let the same mistag class survive one fix already

    What we learned: e94a8cd14 found that Relevant-Rooms can itself be miscomputed (falsely including Originating-Terminal). This event shows a second, independent failure one layer downstream: even when Relevant-Rooms is computed CORRECTLY and explicitly excludes the room in question, the stage that decides which room's lane to score a commit's offPct against still reads Originating-Terminal directly — meaning there are at least two separate call sites in the pipeline that resolve 'which room is this,' and a fix applied to one (or to the field's construction) never reached the other. Patching a value doesn't patch every consumer of that value; that's the structural fact this event adds.

    The missing guard (the deliverable): tests/pmu-simulator/room-attribution-prefers-relevant-rooms.test.mjs — SMOKE: feed the pipeline bb6c10607's real trailers (Originating-Terminal: vault, Relevant-Rooms: builder/operator/laboratory) plus its real changed-file list, assert the room used for offPct/lane scoring is NOT vault; NO SMOKE: feed a commit where Originating-Terminal and the top Relevant-Rooms match genuinely agree, assert the room used is unchanged. Both directions run against real trailer bytes via `git show bb6c10607`, never a hand-typed fixture.

    Detector probe: signal attribution — gzip reef 80% · ballistic walk 20%. Fully saturated (100% lit-to-walked ratio on both sides) with near-identical volume — this reads as SATURATED rather than right-sized: every reachable node the walk touches gets marked lit, which gives zero discriminative signal about lane fit. Concrete adjustment: require a minimum multi-hit or edge-weight threshold before a walked cell counts as 'lit' (e.g. --lit-threshold=2), then re-run this event and the room's other 24 to see whether saturation drops below 100% while the true off-lane signal (the room-attribution bug) stays visible. Experiments: Re-score the last 25 vault out-of-lane events using Relevant-Rooms' best-glob-match instead of Originating-Terminal and measure how many of the 25 stop being out-of-lane entirely — this directly tests whether the mistagged class explains most of the room's recent offPct climb · Log, per receipt, whether the attributed room's owned-surface glob matched ANY changed file; flag zero-match attributions as 'attribution-uncertain' rather than computing offPct against them, and track what fraction of vault's 25 events fall into that bucket · Correlate prewalkOverlap against offPct across the room's last 25 events — if low prewalkOverlap (like this event's 3) predicts high offPct independent of room-attribution correctness, that's a second, real seeding-tightness signal worth its own guard

    Measured outcome (deterministic, not self-graded): room mean offPct —→31 over 4 commits since · dictionary ablation saved 1118B

    A mistagged lane inflates vault's apparent loss ratio with events that were never vault's risk to begin with, which either overprices the room's premium or — worse — trains the underwriting signal to distrust vault's lane data generally; restoring it means the room-attribution fix plus its firing test, so vault's offPct only ever reflects work that actually touched its owned surface.

    Open the signed attestation →

71. 2026-09-04 · 🔒 vault · offPct 25 · mistagged

    fix(gem): the doctor was green over a six-day-dead chain — it audits Safari, the harvest is Chrome

    Why it left the lane: The lane-check machinery tested this commit against 🔒 vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/) purely because Originating-Terminal says 'vault' — but every changed path (scripts/chat/gem-doctor.mjs, scripts/chat/gemini-cdp-scrape.mjs, scripts/ops/initiative-loop.sh, tests/chat/gem-doctor.test.mjs) is scrape-chain/CDP/doctor territory, exactly what the commit's own Relevant-Rooms trailer already names: '🧭 navigator, 🔨 builder, 🧪 laboratory'. Terms like 'chrome-cdp --status', 'signedIn', 'CORPUS_STALE_DAYS', 'gemini profile is signed in' have zero gravitational pull toward patents/legal/data-room — they pull cleanly toward navigator/builder/laboratory, which is exactly where the reality walk landed (offPct=25 relative to vault's lane, which the receipt correctly reports as OUT because vault was never the real target).

    Intent–reality gap: The commit message and its own Relevant-Rooms trailer correctly scope this as a scrape-chain/doctor repair owned by navigator/builder/laboratory; the receipt's lane-check nonetheless ran the offPct test against vault (Originating-Terminal) — the room whose owned-surface glob matches zero of the four changed files — so the 'intent lane' being graded was never the commit's actual claim.

    • grep scripts/pmu/*.mjs and scripts/ops/*.mjs for 'Originating-Terminal' to find every site that reads it as a lane key, not just the one e94a8cd14 already patched
    • add a per-commit debug field to the receipt row: 'laneRoom' (the room actually tested) and 'globMatchCount' for that room vs each Relevant-Room, so a mistag is visible without re-deriving it
    • re-run this sha's receipt with laneRoom forced to 'navigator' (or the union of navigator/builder/laboratory owned-surface) and confirm offPct drops well under 15
    • audit vault's other 25 out-of-lane events in the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms mismatch pattern before assuming they're all genuine vault drift
    • update the .workflow/rooms/vault.html owned-surface doc to state explicitly that Originating-Terminal=vault commits touching scripts/chat/ or scripts/ops/ are NOT vault-lane work

    What we learned: e94a8cd14 fixed the case where Relevant-Rooms falls back to listing a 0-match Originating-Terminal at the top of its own array (measured: room mean offPct 21→12). This event shows a SECOND, independent attribution path: even when Relevant-Rooms is clean and correctly excludes vault entirely (navigator, builder, laboratory — no vault), some other call-site still keys the offPct test off Originating-Terminal directly, bypassing Relevant-Rooms altogether. The room-for-lane-check binding is not centralized in one function — it's read from at least two places, and fixing one left the other live, which is why vault's 14-day count (26 events) hasn't come down despite the prior fix's measured win elsewhere.

    The missing guard (the deliverable): A test that constructs a synthetic commit trailer with Originating-Terminal='vault' (0 owned-surface glob matches against the changed file list) and Relevant-Rooms='navigator, builder, laboratory' (matching 100% of changed files), feeds it through the lane-attribution function, and asserts the resolved laneRoom is one of {navigator, builder, laboratory} — never vault. Should fail loudly today against whichever call-site still reads Originating-Terminal unconditionally.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent = 100% and litReality/walksReality ≈ 100% — full saturation, which for a 4-file, ~150-line diff is a small enough footprint that the walk exhausts its own budget without decaying. That's consistent with the walk being right-sized for this commit's scope rather than under- or over-utilized, but the 100% ratio on both sides is worth watching: a commit this size shouldn't need the full walksIntent=141 budget to converge, and a saturated ratio on every small commit would mask real depth differences. Adjustment: log ply-at-saturation per commit so a genuinely deep/ambiguous commit can be told apart from a small one that merely used its whole budget. Experiments: Re-run this sha's reef seed with the intent corpus built from the union of Relevant-Rooms' (navigator+builder+laboratory) owned-surface docs instead of vault's, and measure whether prewalkOverlap rises from 5 to a value consistent with genuine on-lane work — if it jumps to >30, that's direct evidence the reef seed itself is picking the wrong intent corpus, not just the downstream offPct label · add a per-Relevant-Room offPct computation (test the same reality walk against each of navigator/builder/laboratory's owned-surface glob, not just vault's) and record which room actually resolves offPct<15 for sha bcbbb10fd2 · over the next 20 out-of-lane events, log (Originating-Terminal glob-match-count==0) alongside (offPct>threshold) and measure the correlation — if it's strong, that pairing becomes a cheap pre-filter to auto-flag mistagged events before an intervention pass runs

    Measured outcome (deterministic, not self-graded): room mean offPct —→31 over 4 commits since · dictionary ablation saved 834B

    Testing this commit against vault's lane inflates vault's measured offPct and its 14-day out-of-lane count (26 events) with scrape-infrastructure work it never actually owns, artificially raising that lane's apparent loss-ratio; restoring attribution to the Relevant-Rooms (navigator/builder/laboratory) that the commit itself already named would drop this event's offPct near zero and let vault's premium reflect its real patent/legal surface risk instead of chronic mistagging noise.

    Open the signed attestation →

72. 2026-09-04 · 📐 architect · offPct 19 · structural

    fix(voice): no-false-humility flagged the token, not the amputation

    Why it left the lane: The entire diff lives in tests/voice/no-false-humility.test.mjs — a guard for the voice register — while architect's declared lane (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, CLAUDE.md) owns zero test files. The reef ingest reads path + identifier vocabulary; strings like 'AMPUTATIONS', 'ALLOW', 'blanket disclaimer', 'we make no claim', and the file path itself all seed into the voice-guard region of the lattice, not architect's strategy-doc region, so the walk lit cells the intent lane never claimed.

    The semantic pull: Concrete pull points: the changed identifiers 'AMPUTATIONS' array entries and the removed 'ALLOW' object keyed to 'src/content/blog/2026-08-14-the-physics-of-slop.mdx'; the literal disclaimer vocabulary 'we make no claim(s)'; and the bare file path 'tests/voice/no-false-humility.test.mjs' — none of these tokens or paths intersect architect's owned globs, so the reef seeds and the walk's saturation (litReality 134 == walksReality 134) both land outside the claimed lane with no offsetting anchor inside it.

    Intent–reality gap: The commit message claims this is architect's work ('architect owns the instrument') and names voice/vault as merely Relevant-Rooms, but the receipt classifies lane membership from Originating-Terminal alone against a glob list with no tests/ entry anywhere — so a commit the operator's own trailer convention explicitly sanctions as cross-room reads as 100% off-lane.

    • Ship scripts/ops/room-drift-check.mjs unifying the three identical drafts that never landed (this is the second recurrence of the same unshipped fix)
    • Add 'tests/**/*.test.mjs (guard-instrument maintenance, cross-room)' as a recognized non-content lane in .workflow/rooms/architect.html's owned-surface section, scoped to guard repair not content authorship
    • Add the Relevant-Rooms-union rule as a one-line addition to CLAUDE.md's commit-trailer section (see spec_rewrite)
    • Cross-reference .workflow/rooms/voice.html so it names architect as the sanctioned instrument-fixer for tests/voice/*, keeping register authorship with voice
    • File a stalled-fix flag against ff09774e5 in the delegation ledger — a fix independently spec'd by 3 rooms with 0 shipped instances is a distinct failure worth tracking separately from the drift event itself

    What we learned: A structural fix can be correctly diagnosed AND independently re-derived by three separate rooms and still never ship — convergent diagnosis is not self-executing. ff09774e5 named this exact mechanism and even reported an improved outcome (20→14), but the actual code (room-drift-<room>-check.sh) was never merged, so the same room hit the same class again at offPct 19. The gap isn't detection of the fix, it's the absence of a forcing function that turns '3 rooms proposed the identical patch' into 'someone owns landing it.'

    The missing guard (the deliverable): tests/ops/room-lane-trailer-union.test.mjs: seed a synthetic commit whose Originating-Terminal glob excludes the changed file but whose Relevant-Rooms trailer includes a room whose glob covers it, assert the lane classifier reports IN-LANE (not off-lane) once Relevant-Rooms is unioned in — and assert it reports OFF-LANE today without the fix, so the guard is seen-to-fail per the 2026-08-27 rule before it's trusted.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Saturated, not shallow or under-used: a 1:1 lit-to-walked ratio on both sides means the walk exerts no filtering pressure. Concrete adjustment: replace the binary lit criterion with a ply-decay-weighted threshold so a walked cell only lights when its accumulated weight clears a bar, restoring the walk's ability to discriminate seed noise from real drift instead of just echoing the seed 1:1. Experiments: Re-render this sha's panel with Relevant-Rooms-unioned glob membership (per spec_rewrite) and confirm offPct drops below the 15 threshold for this specific commit · Sample the last 20 guard-fix commits room-tagged similarly (Originating-Terminal ≠ file-domain room) and measure whether prewalkOverlap alone predicts offPct without the walk's contribution, to confirm the 70/30 split generalizes rather than being a one-commit artifact · Ablation the lit-threshold change (binary → decay-weighted) against this sha and 5 neighboring commits offline, comparing driftPct/offPct before and after to quantify whether walk saturation is inflating offPct repo-wide, not just here

    Measured outcome (deterministic, not self-graded): room mean offPct 18→9 over 11 commits since · dictionary ablation saved 675B

    This is sanctioned, trailer-declared cross-room instrument work misread as unbounded scope creep, which inflates architect's apparent loss ratio on paper; premium is restored the moment the detector prices Relevant-Rooms-declared collaboration as in-lane instead of penalizing work the house rules already approved.

    Open the signed attestation →

73. 2026-09-04 · 🎤 voice · offPct 63 · mistagged

    ops(votalai): file the sent DM — 10 sealed predictions with six-percentage grading + the roundup

    Why it left the lane: The commit is filed as 🎤 voice (owned surface: books/tesseract/chapters, blog .mdx, decentralized talk assets), but the diff is entirely operational sales/ops artifacts — a predictions ndjson (data/predictions/2026-09-03-votalai-dm.ndjson) and an HTML sales roundup (docs/ops/votalai-dm-roundup-2026-09-03.html) documenting a LinkedIn DM, a licence webhook fix, and CRM-style prediction rows with adjudicators/falsifiers/p-values. None of these strings — 'Bobby Gupta', 'sealed predictions', 'Nexasure/O'Neill', 'licence webhook hardened', 'stripe_ref' — cluster anywhere near the voice room's lexicon of book/blog/talk content; they cluster with docs/ops/*, data/predictions/*, and tests/license/*, which is 🎩 operator / 🔒 vault territory. The reef placed the mass at those operator/vault coordinates because that is where the vocabulary actually lives, producing offPct=63.

    Intent–reality gap: The commit message claims this is filed FROM room 🎤 voice, but the content is a sales-ops artifact (a sent DM roundup + sealed predictions + a licence webhook incident note) that has nothing to do with voice's owned surface (book chapters, blog posts, decentralized-talk slides) — the work is legitimate and well-scoped, the room label on it is simply wrong.

    • Update `.workflow/rooms/voice.html` (or wherever voice's owned-surface glob lives) to confirm data/predictions/** and docs/ops/** are explicitly NOT in its lane, so future commits from the voice terminal doing ops work get flagged before commit, not after.
    • Add `data/predictions/**` and `docs/ops/**` to the 🎩 operator (or 🔒 vault) room's owned-surface glob if not already covered, so the reef has a correct home to place this mass into.
    • When dictating an ops/sales filing while physically in the voice terminal, either switch terminal first or set `Originating-Terminal` to the room the work actually belongs to rather than the terminal that happened to be open.
    • Cross-reference this event against the room's other 96 out-of-lane events in the last 14 days — if most are this same ops-filed-as-voice pattern, the fix is the glob update above, not a one-off relabel.

    What we learned: This room's persistent 97-event/14-day out-of-lane rate is not one bad commit repeating — it's the terminal-detection convention (Originating-Terminal = physically-open terminal) colliding with content-based lane detection (reef = vocabulary placement): the operator dictates ops filings from whatever terminal is open, and the trailer faithfully records that terminal, but the receipt measures where the WORDS land, not where the FINGERS were — so any session that files ops/sales work while sitting in the voice terminal will structurally mistag every time, and no single commit fix closes it, only a rule that origination-terminal and content-lane are two different axes and the trailer should follow content, not physical terminal.

    The missing guard (the deliverable): A pre-commit check that computes the reef-lane placement of the STAGED diff's semantic content (cheap gzip-NCD reef seed, not the full ballistic walk) and warns when it disagrees with the declared Originating-Terminal room by more than the lane threshold — surfacing the mismatch before commit rather than after the receipt renders.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized, not the problem — litIntent (112) and litReality (115) are nearly equal and walksIntent/walksReality track 1:1, so the walk isn't saturating or truncating asymmetrically (walkTruncated=true is noted but doesn't show up as an imbalance here); the defect is upstream in reef placement/seed overlap, not walk depth. Experiments: Recompute prewalkOverlap for this same commit against the 🎩 operator and 🔒 vault reef seeds instead of 🎤 voice, and confirm overlap jumps well above 3 — if it does, that's direct evidence the mistag (not the walk) drove offPct. · Run the last 20 'ops(votalai)' and 'ops(' scoped commits filed under 🎤 voice through the reef seed step only (skip the full walk) and tabulate their prewalkOverlap distribution vs. commits correctly filed under 🎩 operator — if voice-tagged ops commits cluster at overlap<5 while operator-tagged ones cluster higher, that's the ablation that justifies auto-flagging low-overlap commits pre-commit. · Ablate the reef by adding 'sealed predictions', 'adjudicator', 'falsifier', 'stripe_ref', 'licence webhook' as seed terms under the 🎩 operator lane specifically, then re-run this commit's receipt and check whether offPct drops — confirms whether the fix is a vocabulary/glob one rather than a room-detection one.

    Measured outcome (deterministic, not self-graded): room mean offPct 23→19 over 4 commits since · dictionary ablation saved 835B

    An ops/sales filing mistagged as content-room work inflates that lane's apparent loss-ratio with noise unrelated to its actual craft — the fix is a one-line trailer correction plus a glob update, cheap enough that the lane's premium should reprice back down as soon as the next few 🎤-tagged commits stop being ops notes.

    Open the signed attestation →

74. 2026-09-04 · 🔨 builder · offPct 40 · mistagged

    feat(cta): five angles under a ratchet — Hooper wins on merit, not by being the only one

    Why it left the lane: The 40% off-lane mass traces to two new files that sit outside 🔨 builder's declared owned-surface: scripts/ops/cta-ratchet.mjs (113 new lines) and data/cta-angles.json (49 new lines). Builder's glob is src/app/, src/components/, src/lib/, public/ plus a short named-script allowlist — it contains no scripts/ops/* or data/* entry. The reef's gzip-NCD seed matches vocabulary, not just paths: strings like 'REFUSE', 'AXES', 'THE RATCHET IS min(), NEVER mean()', 'THE FLOOR ONLY RISES' and the file's own comment ('Same shape as scripts/shoot/shoot-ratchet.sh') are the same jargon family as the repo's existing scripts/ops/*.mjs corpus (wip.mjs, rules-for.mjs, quiet-gate.mjs, token-vs-construct.mjs — all cited in CLAUDE.md), which the reef already associates with the operator/ops region of the lattice. That vocabulary and path pull, not any UI-side content, is what dragged litReality to 131 against litIntent's 98.

    The semantic pull: cta-ratchet.mjs's own header explicitly anchors itself to 'scripts/shoot/shoot-ratchet.sh' and reuses the repo's standing ops-governance register ('THE FLOOR ONLY RISES', 'REFUSED rather than silently written', 'inadmissible is not the same as unpopular') — this is the same lexical family the reef has already indexed from a dozen other scripts/ops/*.mjs files, so the seed placement pulls toward that region independent of directory alone.

    Intent–reality gap: The commit message claims one coherent UI feature scoped to the gift-license floater ('five angles under a ratchet'), and the code is in fact correct and coherent — but 40% of its lit mass legitimately lives in scripts/ops/ and data/, paths the room's declared owned-surface never enumerated, so the receipt reads a correct commit as drifted.

    • Add `data/cta-*.json` and `scripts/ops/cta-*.mjs` to 🔨 builder's owned-surface glob in `.workflow/rooms/builder.html`.
    • Add a warn-only pre-commit check comparing every new path in a commit against the Originating-Terminal room's glob, per MISSING GUARD below.
    • Note in scripts/ops/cta-ratchet.mjs's header that it is UI-copy-ranking scoped to GiftLicenseFloater, not general ops infra, so it isn't later claimed by 🎩 operator by directory convention alone.
    • Backfill-check the room's other high-offPct commits (62 events/14 days) for the same scripts/ops+data companion-file pattern before trusting the glob fix as sufficient.
    • Ping 🎩 operator (already named in Relevant-Rooms) confirming scripts/ops/cta-ratchet.mjs is intentionally builder territory, not an accidental drop into their namespace.

    What we learned: Directory-based owned-surface globs undercount a room's real output the moment that room ships a small companion script+data pair alongside a UI change — the room doesn't stop being 'builder' just because the supporting logic lives in scripts/ops/ by repo convention; the glob needs to track WHO the artifact serves, not just WHERE convention places that file type.

    The missing guard (the deliverable): A warn-only pre-commit check (extending hooks/pre-commit) that, for every new file path in the commit, checks it against the Originating-Terminal room's `.workflow/rooms/<room>.html` owned-surface glob AND against every room named in Relevant-Rooms; if a path matches neither, print a WARN naming the room whose glob should probably be widened — this converts a silent 40%-off-lane receipt into an immediate, actionable prompt at commit time instead of a post-hoc intervention three commits later.

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Both sides are fully saturated (litIntent/walksIntent = 100%, litReality/walksReality = 100%) with walkTruncated=true — the walk is hitting its ply/time budget on both sides rather than terminating naturally, so it cannot currently distinguish 'genuinely out-of-lane' from 'ran until the budget ran out.' Concrete adjustment: rerun with a higher --max-depth on this commit alone and check whether litReality keeps growing past 131 — if it plateaus, the walk is right-sized; if it keeps climbing, the budget is currently masking the true spread. Experiments: Re-run prewalkOverlap after stripping repo-wide ops jargon already indexed elsewhere ('ratchet', 'floor rises', 'REFUSE') from the reality corpus for this commit only, to test whether the low 3-cell overlap is genuine novelty or borrowed vocabulary the seed hasn't credited yet. · Since the diff's own comment names its prior art ('Same shape as scripts/shoot/shoot-ratchet.sh'), seed the intent corpus with that referenced file's vocabulary and measure whether prewalkOverlap rises and offPct drops on a re-run. · Ablate data/cta-angles.json out of the reality ingest for this commit and re-measure litReality/offPct, to quantify how much of the 131-cell spread the data file alone contributes versus the script.

    Measured outcome (deterministic, not self-graded): room mean offPct 20→20 over 8 commits since · dictionary ablation saved 922B

    A mistagged-but-correct commit inflates this lane's measured loss ratio without any actual defect, so the fix that restores premium is the glob widening plus the commit-time guard — not a code change — otherwise every future companion-script pattern from this room keeps paying for a boundary the room's own declaration never drew.

    Open the signed attestation →

75. 2026-09-04 · 🎤 voice · offPct 36 · mistagged

    comms(external): drop guard-banned 'byte-identically' from the Bobby draft — outgoing register is shape-identical

    Why it left the lane: The diff's only touched path is docs/comms/external/drafts/2026-09-03-bobby-gupta-the-record-you-did-not-author.md — zero overlap with 🎤 voice's owned globs (books/tesseract/chapters/*.md, src/content/blog/*.mdx, decentralizedtalk/slideData.ts, src/app/6needs/, scripts/voice-*-check.sh). The swapped phrase itself ('recomputes byte-identically' → 'recomputes the same shape') plus surrounding vocabulary ('licensed', 'priced per agent-year', 'CART findings', 'underwriter') is pure comms/external-corpus mass — it pulls the reality walk toward the ☕ network / 📐 architect region of the lattice, not toward book/blog coordinates. prewalkOverlap=5 against walksIntent=106 shows the reef already saw almost no shared seed before the recursive walk even started — the mismatch is baked in at ingest, not manufactured later.

    The semantic pull: 'byte-identically'/'the same shape' plus 'licensed'/'priced per agent-year'/'CART findings'/'underwriter' are outreach-register terms that anchor to the comms/external corpus mass, dragging lit-reality cells toward network/architect coordinates while the claimed intent lane sits on voice's book/blog surface — a one-word swap in a sealed outbound letter, not a syllable of book or blog prose.

    Intent–reality gap: The commit message and trailer correctly name comms(external) work and list ☕ network + 📐 architect as Relevant-Rooms, but Originating-Terminal 🎤 voice is what the receipt scored the lane against — so a commit that never claims book/blog work still gets penalized against a book/blog owned-surface it never touched.

    • Add the Lane-Anchor trailer rule to CLAUDE.md's COMMIT MESSAGE STRUCTURE section
    • Patch the lane-scoring script to prefer changed-file-glob match over Originating-Terminal identity when they diverge
    • Check .workflow/rooms/*.html to confirm docs/comms/external/** is an owned surface of ☕ network (or 📐 architect) — if it is owned by nobody, that is the deeper structural gap
    • Run /recontext to check whether docs/comms/external/ is UNOWNED or CONTESTED across the 9 rooms
    • Add a regression fixture from this exact commit (cce2b3661) to the lane-scoring guard once patched

    What we learned: 🎤 voice's 94 out-of-lane events in 14 days is not per-commit noise — it is a structural consequence of the terminal named 'voice' (Terminal.app) being the physical seat the operator dictates comms/outreach work FROM, while its owned-surface glob is scoped narrowly to book/blog content. Every comms commit authored from that terminal will glob-mismatch until the receipt keys off Relevant-Room content-match instead of Originating-Terminal identity — the prior three lessons for this room recorded outcome numbers with no mechanism because this attribution bug was never named.

    The missing guard (the deliverable): tests/ops/lane-anchor-match.test.mjs: assert that for any commit where Originating-Terminal's owned-surface glob has zero overlap with the changed files, the trailer carries an explicit Lane-Anchor field whose named room's glob DOES match — failing the commit format check (warn-only, per existing convention) rather than silently mis-scoring the receipt.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Both sides are near-saturated (lit:walked ratios ≈1.03 and ≈1.21), not under-lit — but walkTruncated=true on a single-line-diff commit is suspicious: a docsOnly commit this small should never need to hit a depth/time ceiling. Concrete adjustment: raise --max-depth (or the time budget) specifically for docsOnly=true commits and re-render this sha to check whether offPct drops once the walk is allowed to finish. Experiments: Rerun cce2b3661 with walkTruncated forced false (raised max-depth) and diff the resulting offPct against 36 to isolate how much of this drift is truncation artifact vs real mismatch · Regression-test whether prewalkOverlap magnitude predicts offPct across the last 50 commits — if strongly correlated, gate full-walk cost behind a cheap prewalkOverlap pre-check · Seed a distinct docs/comms/**/*.md corpus segment (separate from book/blog) and re-render recent comms-from-voice-terminal commits to check whether they land in-lane once scored against the correct segment

    Measured outcome (deterministic, not self-graded): room mean offPct 21→20 over 4 commits since · dictionary ablation saved 1027B

    A benign, guard-compliant one-word fix got scored as heavy drift purely because lane-attribution keys off physical terminal rather than content ownership, inflating 🎤 voice's false-positive loss-ratio — restored once the receipt anchors the lane to whichever Relevant-Room's owned surface actually matches the changed files.

    Open the signed attestation →

76. 2026-09-04 · 🔨 builder · offPct 42 · mistagged

    feat(initiative): the stencil is placed by the ballistic walk — cascade fan-out feeds impact, the hat rides the brief

    Why it left the lane: None of the 5 changed files (scripts/ops/initiative-frame.mjs, scripts/ops/initiative-score.mjs, scripts/ops/initiative-stencil.mjs, scripts/ops/stencil-ab.mjs, tests/ops/initiative-stencil-walk.test.mjs) fall inside 🔨 builder's declared owned surface (src/app/, src/components/, src/lib/, public/, build-book-html.sh, convert-chapters-to-html.mjs, hooks/pre-commit, voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json) — zero overlap. The diff's own vocabulary is PMU/architecture-flavored: 'pickStencilWalk', 'walkShape from unified-drift', 'cascade heat read at each reef domain's own coordinate', 'gzip-NCD lit score', 'shortlex-names.mjs', 'PMU_BINARY forced absent' — none of it shares a token with builder's UI/component surface. That vocabulary mismatch is exactly why prewalkOverlap sits at 4: the reef seed found almost nothing in builder's domain to anchor on, so the walk (already near-saturated at litIntent=127/144, litReality=132/144, walkTruncated=true) spread into whichever reef domains the initiative/cascade/stencil words actually live at — coordinates outside anything builder's glob recognizes as in-lane.

    The semantic pull: 'pickStencilWalk', 'walkShape', 'cascade', 'reef domain', 'ShortLex', 'unified-drift.mjs' pull toward the PMU/architecture coordinate cluster; scripts/ops/ has no owned-surface glob at all in the room-ownership config, so this pull lands nowhere builder (or apparently any room) legitimately claims — the receipt then measures the whole thing as builder's overflow because builder's terminal happened to originate the commit.

    Intent–reality gap: The commit message (and Relevant-Rooms trailer) claims 🔨 builder + 🧪 laboratory + 📐 architect jointly own this work, but the receipt measured 0 of 5 changed files against builder's actual owned-surface glob — the room-attribution is asserted, not derived from the files that moved.

    • Add scripts/ops/** and tests/ops/** to 📐 architect's 'Owned file/dir surface' in .workflow/rooms/architect.html — this is the actual missing lane.
    • Re-run the lens-reef room-attribution script over the last 30 days of scripts/ops/ commits and check whether their offPct drops once ownership is assigned instead of defaulted to Originating-Terminal.
    • Add a CLAUDE.md line under DELEGATE-TO-SUBTASKS or ROOM SURFACE noting scripts/ops/ + tests/ops/ as architect-owned, so future commits don't repeat the same default-to-terminal mistagging.
    • Cross-check tests/ops/ and scripts/ops/ against 🧪 laboratory's owned surface too — stencil-ab.mjs is an A/B measurement harness, which is squarely laboratory territory; ownership may need to be split rather than assigned wholesale to architect.
    • Audit .workflow/rooms/*.html for any other directory with zero glob coverage across all 9 rooms — scripts/ops/ is unlikely to be the only orphan.

    What we learned: The prior three lessons for this room were facts-only (no sensemaking cached), so this is the first mechanism-level finding: 🔨 builder's out-of-lane events are not (at least this time) the room doing cross-domain work — they can be the room-ownership glob having a gap that scripts/ops/ (and possibly other utility directories) falls into, so ANY terminal that happens to commit there gets blamed for drift the config never gave it a home for.

    The missing guard (the deliverable): A pre-commit or reef-seed check that flags 'orphan surface': if every changed file in a commit matches ZERO room's owned-surface glob, the drift receipt should mark the commit UNOWNED and skip lane-threshold scoring against whichever terminal originated it, rather than silently defaulting attribution to Originating-Terminal and then penalizing that room's offPct trend.

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. litIntent(127/144≈88%) and litReality(132/144≈92%) are both close to the 144-cell ceiling, and walkTruncated=true is set — this is edging toward the SATURATED regime the anti-rules ledger warns about (|shape|/144 degenerating content-independent), where the walk's signal stops discriminating. Concrete adjustment: apply the same saturation discount this very commit just shipped for initiative-score.mjs (credit nothing from a saturated cascade) to the drift-receipt's own walk — flag rows where litIntent or litReality exceeds ~80% of 144 as 'saturated, offPct unreliable' rather than scoring them at face value. Experiments: Backtest: for the last 20 receipts room-wide, compute correlation between prewalkOverlap and offPct — if low prewalkOverlap (<10) reliably predicts high offPct (>30), surface a warning at commit time before the walk runs rather than after. · Re-run this exact commit's receipt after adding scripts/ops/** to architect's owned-surface glob, and measure whether offPct drops for the SAME sha (glob-attribution is deterministic and cheap to re-derive without re-running the walk). · Add a saturation flag to the receipt row itself (litIntent/144 and litReality/144 as explicit percentages) so 'near-saturated' rows are visible in the ledger without re-deriving them from raw walk counts each time.

    Measured outcome (deterministic, not self-graded): room mean offPct 20→19 over 6 commits since · dictionary ablation saved 1050B

    An out-of-lane event caused by a missing ownership glob (not by careless cross-domain work) is a config-debt loss, not a code-quality loss — premium on this lane is restored the moment scripts/ops/** gets a real owner, cheaper than any code-side fix would be.

    Open the signed attestation →

77. 2026-09-04 · 🔨 builder · offPct 38

    feat(iamfim): both discs back at the foot, expandable — and the gift floater moves top-right

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 21→26 over 6 commits since

    Open the signed attestation →

78. 2026-09-04 · 🔨 builder · offPct 21 · mistagged

    feat(rules): the explanation is the sufficient statistic for approval

    Why it left the lane: The diff touches exactly two files, CLAUDE.md (root-level rules doc) and tests/ops/explanation-is-sufficient.test.mjs, and neither path matches any entry in 🔨 builder's owned-surface list (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The diff's own vocabulary — 'AXIOM 1's W6', 'sufficiency contract', 'ENFORCEMENT: partly mechanical', 'the irrevocable', 'ratchet' — is governance/meta-rule language, which is exactly the register the reef routes toward vault/operator coordinates, not toward builder's UI/app/script surface. The commit's own trailer even names the room explicitly: 'Vault is the QC room because this rule governs what may be done without asking.' The reef placed the reality mass where the content actually lives; the lane claim (builder) never covered that territory to begin with.

    The semantic pull: Strings pulling the placement off builder's lane: 'THE EXPLANATION IS THE SUFFICIENT STATISTIC FOR APPROVAL', 'HARD RULE, operator-designated', 'ENFORCEMENT: partly mechanical, and the split is stated rather than papered over', and the test file's own header block ('THE COMMIT MESSAGE IS THAT TEXT', 'a scorer... would be theatre'). None of these terms co-occur with src/app|components|lib vocabulary anywhere in the diff — there is zero UI/build/component surface in this commit at all, so 100% of the changed bytes sit outside builder's glob and the receipt's 21% offPct is actually an undercount of that mismatch.

    Intent–reality gap: The commit claims Originating-Terminal 🔨 builder, which the pipeline's lane check maps to builder's owned code surface; the receipt measured reality landing in CLAUDE.md governance prose and an ops test, a domain the commit's own Relevant-Rooms trailer already assigns to 🎩 operator / 🔒 vault first. The claim and the measurement were never going to agree — the commit message is honest about where the work belongs and the room-attribution field is what's wrong.

    • Add CLAUDE.md and tests/ops/** to vault's (or operator's) owned-surface glob in data/rooms.json / .workflow/rooms/vault.html so future governance commits route to the correct lane regardless of which terminal ran them.
    • Patch the lane-detection step in the receipt pipeline to fall back to Relevant-Rooms[0] when Originating-Terminal's owned-surface set has zero file overlap with the diff.
    • Backfill-check the other 47 builder out-of-lane events in the last 14 days for the same signature (Originating-Terminal builder, zero files in builder's glob) — if most match, this is the dominant cause of that room's chronic offPct, not 48 independent incidents.
    • Add a guard asserting the fallback fires: a commit whose Originating-Terminal owns none of the changed files must be graded against Relevant-Rooms[0]'s glob, verified against this exact commit as the fixture.

    What we learned: CLAUDE.md's own COMMIT MESSAGE STRUCTURE section explicitly licenses Originating-Terminal and file-relevance to diverge ('that's fine and worth recording') — but the drift-receipt pipeline still uses Originating-Terminal as the sole lane key. That documented, sanctioned divergence is a standing guarantee that any terminal used for miscellaneous cross-domain work (governance edits, ops tests, anything not tied to a specific room's surface) will accumulate false-positive out-of-lane events forever, independent of the actual quality or correctness of the work — 48 events in 14 days for one room is the accumulated cost of a permission the file itself grants but the detector never learned to read.

    The missing guard (the deliverable): tests/ops/lane-routing-falls-back-to-relevant-rooms.test.mjs — asserts that when a commit's Originating-Terminal owns zero of the changed files (computed from the room's real owned-surface globs, not a string match), the drift-receipt's intent lane is Relevant-Rooms[0], with this commit (d33b1181c) as the seeded fixture: Originating-Terminal=builder, changed files=[CLAUDE.md, tests/ops/explanation-is-sufficient.test.mjs], expected lane=vault.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent = 140/140 and litReality/walksReality = 138/138 — a 100% hit rate on both sides, combined with walkTruncated=true. That combination means the walk hit its depth/budget ceiling while every step was still finding matches, i.e. it never reached the natural boundary where non-matching cells stop lighting — it was cut off mid-saturation rather than converging. Adjustment: raise --max-depth (or the walk's step budget) for commits with large prose diffs like this one (a multi-paragraph HARD RULE addition), and check whether litIntent/litReality drop below 100% once given room to differentiate. Experiments: Re-run this commit's reef seed placement with CLAUDE.md/tests/ops/** added to vault's owned-surface glob and measure whether prewalkOverlap rises above 12/140 when the correct lane is used. · Increase --max-depth for commits whose diff exceeds a size threshold (this one adds ~1KB of prose) and measure whether litIntent/litReality fall below the 140/140, 138/138 saturation seen here. · Backfill the last 48 builder out-of-lane events with the Originating-Terminal-vs-changed-files overlap check and report what fraction share this commit's zero-overlap signature.

    Measured outcome (deterministic, not self-graded): room mean offPct 17→14 over 10 commits since · dictionary ablation saved 1209B

    An event that traces entirely to a mis-keyed lane rather than misplaced work should not raise builder's loss ratio at all — this commit's premium impact is restored the moment lane-routing uses Relevant-Rooms[0] for zero-overlap commits, since the underlying work (a HARD RULE + its guard) is exactly where it belongs.

    Open the signed attestation →

79. 2026-09-04 · 🔨 builder · offPct 16

    feat(initiative): phase 1 — the miner finds spans and cannot act on one

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 17→14 over 10 commits since

    Open the signed attestation →

80. 2026-09-04 · 🎤 voice · offPct 30 · structural

    feat(predictions): seal the 2026-09-02 forecasts and ship the half that settles them

    Why it left the lane: The new script's own comments and the commit message are written in the repo's mandated paradox/aphoristic voice — 'A forecast nobody settles is not a forecast, it is a mood', 'an uncounted claim reads exactly like a settled one', 'IT REFUSES TO BE BLIND', 'Sixty rows into the register opened this morning, in its schema rather than a rival one' — prose that is lexically indistinguishable from books/tesseract/chapters/*.md and src/content/blog/*.mdx, which IS 🎤voice's owned lane. The gzip-NCD reef seed (project_canonical-sensor-is-gzip-ncd-not-simhash) measures text-compression similarity, so a commit whose PROSE reads like voice's manuscript pulls intent-mass toward voice-adjacent coordinates regardless of which files actually changed — and the changed files (data/predictions/2026-09-03-boriscap.ndjson, scripts/predictions/overdue.mjs, tests/predictions/register-is-append-only.test.mjs) sit in none of voice's owned globs. prewalkOverlap=4 against litIntent=litReality=134 shows the two lanes seeded almost entirely disjoint before the walk even ran.

    The semantic pull: Literary/aphoristic sentence construction in both the commit message and the code comments ('a resolution with no evidence is an opinion', 'a forecast with no date is a belief', 'SUFFICIENT FOR:' / 'NOT SUFFICIENT FOR:') is the exact register the reef associates with voice's blog/book corpus — the confound is style, not subject matter.

    Intent–reality gap: The commit message frames this as sealing a forecast register and shipping its resolver — engineering infrastructure the commit's own Relevant-Rooms trailer correctly assigns to 🎩operator/📐architect/🧭navigator — but because the prose describing it is written in the CLAUDE.md-mandated literary house voice, the receipt (keyed to Originating-Terminal: 🎤voice) measured 30% of reality mass landing outside that lane, double the 15% threshold.

    • Add scripts/predictions/**, tests/predictions/**, data/predictions/** to 📐architect's (or 🧭navigator's) owned-surface glob in .workflow/rooms/*.html — currently no room's owned surface covers this new directory at all
    • Patch the receipt's room-resolution step to prefer Relevant-Rooms over Originating-Terminal when the changed files match none of Originating-Terminal's globs
    • Add the CLAUDE.md clarifying sentence above to COMMIT MESSAGE STRUCTURE
    • Sample the room's other 88 out-of-lane events from the last 14 days and check what fraction share this same Originating-Terminal≠Relevant-Rooms[0] pattern — this single event may be the tip of a much larger structural count
    • Guard: tests/pmu-simulator/lane-uses-relevant-rooms.test.mjs asserting the fallback fires on a fixture shaped like this commit

    What we learned: CLAUDE.md's own trailer format already carries the fix (Originating-Terminal vs Relevant-Rooms are explicitly two different axes) but the drift-receipt pipeline apparently only consumes one of them for lane measurement — the spec was right and the code never read the field that would have caught this, which is a different failure than 'no rule existed.'

    The missing guard (the deliverable): tests/pmu-simulator/lane-uses-relevant-rooms.test.mjs — feed the pipeline a fixture with Originating-Terminal=voice, changed files under scripts/predictions/**, and Relevant-Rooms=[architect,...]; assert the computed offPct is measured against architect's glob, not voice's, and stays under threshold once the correct lane is used.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Right-sized in breadth (134=134, neither side starved) but walkTruncated=true means it stopped short of full depth under the time budget. Since the seeds already diverged at ply 0 (prewalkOverlap=4), more depth would not reconcile a lexical-style confound — the walk is being asked to fix a seeding problem it structurally cannot fix. Adjustment: don't raise --max-depth here; instead re-seed the reality corpus with comment-prose stripped and check whether prewalkOverlap rises, which would confirm the confound lives at the seed stage. Experiments: Re-run this commit's ingest with the reality corpus limited to identifiers + non-comment code lines (strip the paradox-voice comment blocks) and measure whether offPct drops below 15 — isolates whether comment-prose vocabulary is the actual driver · Sample the room's other 88 out-of-lane events from the last 14 days and check what fraction have Relevant-Rooms trailer disjoint from Originating-Terminal — confirms whether this mechanism explains most of the room's drift or just this one commit · Compute prewalkOverlap distribution split by 'Originating-Terminal == Relevant-Rooms[0]' vs '≠'; if the mismatched group has systematically lower prewalkOverlap, that becomes a cheap pre-walk early-warning signal, computable before spending the full walk time budget

    Measured outcome (deterministic, not self-graded): room mean offPct 22→19 over 7 commits since · dictionary ablation saved 993B

    A structural confound that lets voice's lane absorb any commit written in the house literary voice inflates 🎤voice's measured loss-ratio with work it never touched while understating 📐architect's real load; restoring the premium means re-keying lane attribution to Relevant-Rooms so each room's offPct reflects only the surface it actually owns.

    Open the signed attestation →

81. 2026-09-04 · 🔨 builder · offPct 21 · mistagged

    test(gem-loop): keep wo1's firing tests, drop my duplicate, add the static companion

    Why it left the lane: The claimed lane is `tests/ops/gem-loop.test.mjs` doing test-suite housekeeping (drop a duplicate, keep wo1's version, add a companion). But the diff's actual mass is a DELETION of 46 lines carrying the semantic payload — `execFileSync`, `expandsCleanly(assignLine, useLine)`, `SMOKE:` / `NO SMOKE:` test bodies that spawn bash 3.2 and assert on `FRAME_ARGS` — plus a `readFileSync`/regex-extraction block that read `scripts/ops/gem-loop.sh` directly. Removing the two firing tests (not just merging a duplicate) and rewriting the remaining static test to read `SH` (a module-level fixture, not a local `readFileSync`) pulls the placement toward `scripts/ops/gem-loop.sh`'s own coordinate and toward whatever shared-fixture/setup code now defines `SH` — content that lives structurally closer to the builder-owned script surface than to a narrow 'test bookkeeping' claim. The commit message frames this as subtraction (drop duplicate, keep the better test) but the diff is actually a rewrite of the remaining test's plumbing (`SH` replacing `readFileSync(join(REPO, ...))` inline), which is new reality mass the message never names.

    The semantic pull: Strings `execFileSync`, `expandsCleanly`, `FRAME_ARGS`, `set -uo pipefail`, and the shift from inline `readFileSync(join(REPO, 'scripts/ops/gem-loop.sh'))` to a bare `SH` symbol are the pull — they anchor the remaining test's identity to the shell script's own bytes and to a shared fixture setup, not to the 'test file cleanup' story the message tells.

    Intent–reality gap: The message claims a narrow subtractive act — 'drop my duplicate, add the static companion' — but the diff also silently changes the surviving test's fixture access pattern (inline `readFileSync`+`join(REPO,...)` → a shared `SH` constant), which is a structural edit to shared test scaffolding the commit narrative never mentions.

    • Confirm `SH` is defined once near the top of tests/ops/gem-loop.test.mjs (e.g. `const SH = readFileSync(join(REPO,'scripts/ops/gem-loop.sh'),'utf8')`) and that no other test in the file still does its own ad-hoc `readFileSync` of the same path — consolidate to prevent silent drift between fixtures.
    • Add one line to the commit-message convention doc noting that a 'drop duplicate' housekeeping commit must call out any accompanying fixture/plumbing change, even a small one.
    • Verify wo1's kept firing tests and the new static test share the exact same `SH` read (byte-for-byte) so a future edit to gem-loop.sh can't pass one test and silently fail the other from stale fixture state.
    • Add a short code comment above the `SH` constant declaration: 'shared by all tests in this file — edit once, both firing and static checks track it.'

    What we learned: A commit that narrates itself as pure subtraction ('drop duplicate, keep theirs') can still carry a real edit to shared scaffolding (the `readFileSync` → `SH` fixture consolidation) that the message never names — 'removal' commits are not exempt from the intent/reality gap check just because the diffstat is net-negative.

    The missing guard (the deliverable): A pre-commit check on test files that flags when a commit's diff introduces or removes a *new top-level fixture symbol* (a `const X = readFileSync(...)` or similar module-scope binding) without that symbol's name appearing anywhere in the commit message — forcing 'I changed how tests read fixtures' to be stated explicitly rather than absorbed into an unrelated headline claim.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Right-sized on spread (124 vs 125 walks, near parity) but the low prewalkOverlap=4 suggests the walk is being asked to do too much correction work from a weak initial seed; concrete adjustment: increase reef seed weight for commits touching test-fixture-only files (a class already distinguishable by path glob) so intent doesn't have to be walked as far to reach the fixture-plumbing coordinates. Experiments: Recompress the last 20 gem-loop.test.mjs-touching commits with a seed dictionary boosted for 'readFileSync', 'execFileSync', 'FRAME_ARGS' and measure whether prewalkOverlap rises above 4 for this commit class. · Run the walk with --max-depth reduced by 1 on commits whose diffstat is >70% deletion (like this one, 46 of 53 lines removed) and compare offPct — test whether deletion-heavy commits systematically over-walk into reality-side noise. · A/B the reef seed with and without the diff's fixture-symbol names (SH, expandsCleanly) included in the corpus ingest, measuring whether offPct on this commit class drops below the 15 threshold.

    Measured outcome (deterministic, not self-graded): room mean offPct 17→20 over 6 commits since · dictionary ablation saved 897B

    A mistagged-but-benign drift on an owned, low-risk test file nudges this lane's loss ratio only slightly; premium is restored by the spec_rewrite (naming the fixture change explicitly) and the missing_guard (warn-only fixture-symbol check), not by any code change since the reality-side code was correct.

    Open the signed attestation →

82. 2026-09-04 · 🔨 builder · offPct 16 · structural

    feat(initiative): declare the end-state — direction beats proximity on the backward chain

    Why it left the lane: The diff's entire semantic mass sits in scripts/ops/initiative-pull.mjs (endState(), chainRung(), pullStrength()), data/end-state.json, and tests/ops/initiative-pull.test.mjs — three paths, none matching any of builder's owned-surface globs (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, scripts/voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json). The commit carries 'Originating-Terminal: 🔨 builder' and the receipt's room-for-lane-check evidently keys off that trailer rather than off 'Relevant-Rooms: 🎩 operator, 🔨 builder, 📐 architect' (operator listed first) or off the changed-file set itself — so terms like 'chainRung', 'endState', 'milestone', 'pullStrength' dragged reality into an ops/strategy coordinate the builder lane was never fenced to cover.

    The semantic pull: The strings 'scripts/ops/initiative-pull.mjs', 'chainRung', 'data/end-state.json', 'tests/ops/initiative-pull.test.mjs' carry zero token overlap with builder's fenced vocabulary (app/, components/, lib/, public/, build-book-html, convert-chapters-to-html, voice-*-check, decentralizedtalk, thetacoach-crm-mcp, vercel.json) — the reef's gzip-NCD seed found almost nothing to grip in-lane (prewalkOverlap=7 against walksIntent=131), so the walk that followed had no honest in-lane material to expand from and spread the full 125-cell reality walk into out-of-lane ops/strategy territory instead.

    Intent–reality gap: The commit message claims a strategy/ops deliverable — a backward-chain end-state and a re-weighted pull metric — filed under a builder-terminal trailer; the receipt measured 125 lit reality cells with 16% of that mass sitting outside whatever lane 'builder' fences, because none of the three changed files (scripts/ops/, data/, tests/ops/) were ever inside that fence to begin with.

    • Add scripts/ops/**, data/*.json, tests/ops/** to the operator (or architect) room's owned-surface glob in .workflow/rooms/*.html so ops-pipeline commits get a real fence to check against
    • Locate and patch the receipt's room-selection code to prefer file-glob ownership / first Relevant-Rooms entry over Originating-Terminal
    • Audit the 58 builder-room out-of-lane events in the last 14 days for the same defect signature (changed files intersecting zero builder-owned paths)
    • Add the CLAUDE.md sentence quoted in spec_rewrite
    • Add the guard test named in missing_guard
    • Re-render trip-encircled-e73455679f.png after the config fix to confirm offPct drops

    What we learned: The room attached to a drift-receipt event is sourced from Originating-Terminal (physical terminal), not from computed file-glob ownership or Relevant-Rooms — so ops/architecture work typed from whichever terminal happens to be open gets fence-checked against that terminal's unrelated owned-surface and reads as 'out of lane' even when the code is perfectly placed under scripts/ops/ and tests/ops/. This is new relative to the prior three facts-only interventions for this room, which never compared the diff's files against the owned-surface glob and whose measured outcomes (room mean rising 16→46, 15→16) suggest they were patching symptoms of this same unmeasured room-selection defect.

    The missing guard (the deliverable): tests/ops/room-attribution-matches-surface.test.mjs — a FIRING test asserting that for any commit, the room used by the drift-receipt's lane fence-check equals the room whose .workflow/rooms/*.html owned-surface glob covers the majority of the commit's changed files (falling back to the first Relevant-Rooms entry on a tie/miss), never the raw Originating-Terminal trailer when it diverges from both. SMOKE: replay this exact commit's file set (scripts/ops/initiative-pull.mjs, data/end-state.json, tests/ops/initiative-pull.test.mjs) tagged Originating-Terminal=builder and assert the guard flags the mismatch. NO SMOKE: replay a genuine builder commit touching only src/components/*.tsx and assert it stays quiet.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true means the walk hit its depth/time budget before natural termination on a 3-file, cross-directory commit (scripts/ops/, data/, tests/ops/) — right-sized for a single-lane commit but likely cut short here given the wider surface. Adjustment: rerun this sha with --max-depth bumped one notch and diff offPct against this run; if offPct moves, the truncation is inflating/deflating the number and the budget needs a floor keyed to files-touched-count rather than a flat default. Experiments: Add scripts/ops/**, data/*.json, tests/ops/** to the operator room's owned-surface glob, then re-render trip-encircled-e73455679f.png for this sha and measure whether offPct drops toward 0 — a direct ablation of the room-attribution hypothesis · Reseed prewalkOverlap using the first Relevant-Rooms entry (🎩 operator) instead of Originating-Terminal for this sha and compare prewalkOverlap (currently 7) — if it rises materially, seed-selection is confirmed as the primary driver, not the walk · Audit the 58 builder out-of-lane events over the last 14 days: compute file-glob overlap with builder's owned surface for each, and report what fraction (like this one) touch zero builder-owned paths — a measurable test of whether the prior intervention's regression (16→46) is a room-selection artifact rather than a builder-discipline problem

    Measured outcome (deterministic, not self-graded): room mean offPct 19→22 over 4 commits since · dictionary ablation saved 601B

    This event is a false-positive premium tick against the builder lane — the code itself (a well-tested backward-chain ranking change) is not the risk; the risk is that dozens of similar events are inflating builder's measured loss-ratio with ops-pipeline work it was never fenced to cover, and premium only normalizes once scripts/ops/**, data/*.json, tests/ops/** get a real owning room and the receipt stops keying lane-checks off Originating-Terminal.

    Open the signed attestation →

83. 2026-09-04 · 🔨 builder · offPct 19 · structural

    feat(tape): every commit seals to the tape — the write-lock had never been called

    Why it left the lane: No region narrative was cached for this sha, so the coordinate-level story isn't available — but the deterministic receipt row plus the diff stat already carries the mechanism. The diff is 100% scripts/pmu/tape-on-commit.mjs, tests/pmu-simulator/tape-on-commit.test.mjs, and hooks/post-commit. None of those three paths appear in 🔨 builder's owned surface (src/app/, src/components/, src/lib/, public/, scripts/build-book-html.sh, scripts/convert-chapters-to-html.mjs, hooks/pre-commit, .git/hooks/pre-commit, scripts/voice-*-check.sh, decentralizedtalk/, thetacoach-crm-mcp/, vercel.json — note hooks/pre-commit is owned, hooks/post-commit is not). The vocabulary that dragged the walk elsewhere is dense PMU jargon absent from builder's surface entirely: 'tape-walk-worker', 'write-lock', 'PENDING_WALK', 'cursor_id', 'walkShape', 'unified-drift.mjs', 'pmu-onchip', 'the unidirectional contract'. prewalkOverlap=5 is low relative to litIntent/litReality of 133/141 — the reef seed barely agreed before the walk ran, then the walk (walkTruncated=true) saturated both sides near the lattice ceiling. The commit's own trailers already flag the mismatch honestly: Originating-Terminal is 🔨 builder (the physical terminal), but Relevant-Rooms lists 🔨 builder, 🧪 laboratory, 🔒 vault — the author knew this was cross-room work.

    Intent–reality gap: The commit claims to be sealing commits to the tape (infra wiring, correctly scoped and even self-documented as cross-room via the Relevant-Rooms trailer), while the receipt measured it against 🔨 builder's lane — a lane whose owned-file-surface has zero overlap with any of the three changed files. The gap isn't that the work drifted; it's that the lane the receipt checked against could never have contained this work.

    • grep scripts/pmu for where Originating-Terminal is parsed for lane/offPct scoring and confirm it currently ignores Relevant-Rooms
    • confirm hooks/post-commit, scripts/pmu/tape-*.mjs, and tests/pmu-simulator/tape-*.test.mjs are actually claimed by 🧪 laboratory or 🔒 vault's owned-surface glob in .workflow/rooms/*.html — add them if missing
    • add a pre-commit WARN (not block) that computes Originating-Terminal's glob-overlap with staged files and prints the top-scoring Relevant-Room by comparison when overlap is 0%, so the trailer is self-correcting at compose time
    • backfill 🔨 builder's last 14 days of out-of-lane events against a 0%-overlap check to see what fraction of the room's 59 events are this same structural mismatch versus genuine content drift
    • add the owned-surface-overlap percentage as a field on the deterministic receipt row itself, next to offPct, so future lessons can separate 'lane could never fit' from 'lane fit and still drifted' without re-deriving it

    What we learned: The prior three lessons for this room were facts-only (no sensemaking ran), so nothing named the mechanism before now: Originating-Terminal and the receipt's lane-threshold check are the same field today, even though CLAUDE.md explicitly designed Originating-Terminal to legitimately diverge from file-relevance ('that's fine and worth recording'). Every honestly-labeled cross-room commit made from the builder terminal — which this repo's own convention invites via the Relevant-Rooms trailer — is therefore scored against a lane it structurally cannot occupy, independent of whether the work itself was correct or well-scoped.

    The missing guard (the deliverable): A room-surface-coverage assertion: for any commit, at least one room in {Originating-Terminal} ∪ {Relevant-Rooms} must own a nonzero share of the changed files, and the receipt's lane-threshold check must use that covering room rather than Originating-Terminal unconditionally. Without it, a commit can be perfectly labeled (as this one was, naming laboratory and vault) and still fail a threshold that no room-selection choice could have passed.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow: litIntent 133 and litReality 141 are both close to the 144-cell ceiling and walkTruncated=true confirms the walk hit its time/depth budget before resolving cleanly. Concrete adjustment: for commits this vocabulary-dense, lower --max-depth or add a convergence early-stop once new lit cells per ply drops below a small threshold, so the walk reports a clean (non-truncated) coverage number instead of a truncated near-saturation reading that adds noise to offPct run-to-run. Experiments: Re-run this commit's receipt with --max-depth reduced ~30% and confirm offPct/driftPct hold within a few points of 19/1.046 — if stable, the truncation isn't adding real signal for PMU-heavy commits and the walk can be shortened without changing verdicts. · Log the owned-surface overlap percentage on the receipt row (see missing_guard) and re-run the last 14 days of 🔨 builder's 59 out-of-lane events against it, to measure what fraction are this same 0%-overlap structural case versus genuine drift. · Ablation: seed the reef with Relevant-Rooms' owned-surface vocabulary (not just Originating-Terminal's) for commits whose trailer lists more than one Relevant-Room, and measure whether prewalkOverlap rises above 5 — if it does, this event's low-overlap reading is a seeding artifact, not real semantic distance.

    Measured outcome (deterministic, not self-graded): room mean offPct 20→18 over 4 commits since · dictionary ablation saved 801B

    Taken alone this is a mislabeled-lane false positive, not realized loss — the underlying risk (unbounded semantic wander) didn't occur — but it inflates 🔨 builder's apparent 59-event/14-day loss rate with noise from correctly-labeled cross-room work; premium on this lane is restored by fixing lane-resolution to score cross-room commits against a room that can structurally contain them, not by tightening the walk itself.

    Open the signed attestation →

84. 2026-09-03 · 🎩 operator · offPct 32 · mistagged

    post(substrate): the two things the banking analogy hides — which λ, and the matrix/seed inversion

    Why it left the lane: The diff touches only public/exposition/the-third-address-space.html (a rendered exposition page, docs-only per the receipt), while the commit message is dense with operations/pipeline-adjacent vocabulary — 'lattice', 'walk', 'ordering', 'propagation', 'matrix/seed inversion', 'ledger', 'crossing' — terms that gzip-NCD grips against B3.Tactics.Signal (11x) and C1/C2/C3 Operations.Grid/Loop/Flow (6/4/4x) because those coordinates are LEXICALLY DEFINED by exactly this vocabulary (grid/loop/flow/signal are literal reef anchor words), even though the actual bytes changed are prose edits inside one static HTML file with no code, no script, no pipeline logic touched.

    The semantic pull: Real strings from the diff/message pulling mass toward the encircled coords: 'buildIntentGrid' and 'j > i' (→ C1.Operations.Grid), 'the walk is directed', 'ballistic cascade'-adjacent 'nilpotent...finite polynomial' (→ C2.Operations.Loop / recursion), 'S = 2.43 read as effective depth', 'mean reach r/(1-r)' (→ C3.Operations.Flow, throughput/depth language), and 'k_E = 0.003 per crossing', 'the crossing is booked twice' (→ B3.Tactics.Signal, since 'crossing'/'signal'/'ledger entries are addresses' is this room's own defining vocabulary). None of these are code operations — they are MATH PROSE describing the architecture in the register the architecture's own code comments use, so the reef mistakes description-of-the-machine for operation-of-the-machine.

    Intent–reality gap: The commit message stakes a claim entirely in Operator/exposition territory ('second-pass math review of the crash course', a docs-only content edit to one HTML page) and the receipt's own docsOnly=true flag agrees — yet 100% of the lit reality mass fell in Tactics/Operations coordinates the commit never claims to touch, because the PROSE is ABOUT those coordinates without being WORK IN them.

    • Add a 'exposition-only' commit-trailer convention (e.g. `Content-Scope: exposition-prose-only`) that the receipt pipeline can read as a lane-narrowing hint for docsOnly=true commits.
    • Extend the ingest to down-weight math/architecture vocabulary specifically inside public/exposition/*.html, since that directory is inherently ABOUT the machine, not code touching it — a standing lexical discount.
    • Cross-reference the room's owned-file-surface list (docs/03-product, data/trust-debt/, etc.) against public/exposition/ — it currently owns none of the exposition tree, so any exposition edit is structurally guaranteed to look homeless.
    • File-level: register public/exposition/ in .workflow/rooms/*.html as an owned-or-shared surface for 🎩 operator so future placements have a legitimate home lane to compare against instead of defaulting fully off-lane.

    What we learned: A docs-only page that EXPLAINS the pipeline's math in the pipeline's own jargon is structurally indistinguishable, to a gzip-NCD reef, from code that IMPLEMENTS that math — the detector grips on shared vocabulary, not on the AST/file-type boundary the docsOnly flag already knows about. The docsOnly=true signal is currently informational only; it is not yet fed back into the lane-scoring weight.

    The missing guard (the deliverable): A test asserting that when a commit's receipt row has docsOnly=true AND changes zero files under any room's registered code-surface globs, the encircled-region weighting for that commit is capped/discounted by a docs-only factor before offPct is computed — i.e. docsOnly should demonstrably soften the lane penalty, not just be logged as a flag nobody reads.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent=120 vs litReality=101 with fill 0.98% (well under the 70% ceiling) says the walk is right-sized in breadth, not saturated or under-filled — the problem is upstream of the walk, in what got seeded. Adjustment: none needed to walk depth/breadth; the fix belongs in seed weighting (discount math-vocabulary seeds inside public/exposition/ paths), not in walk parameters. Experiments: Run the same commit's receipt with a synthetic version of the message that strips all math-jargon nouns (lattice/matrix/ledger/crossing/signal) but keeps the docs-only claim, and measure whether offPct drops below 15 — isolates whether vocabulary alone drives the miss. · Add a docsOnly-weighted seed multiplier (e.g. 0.5x) in the gzip-NCD seed stage specifically for public/exposition/*.html and public/pmu-demo/ paths, then re-run the last 13 operator-room out-of-lane events and measure the new mean offPct. · A/B the prewalkOverlap threshold: commits with prewalkOverlap<=2 and docsOnly=true are flagged for manual room-surface review before offPct is finalized — measure how many of the room's 13 recent events this would have caught.

    Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 6 commits since · dictionary ablation saved 961B

    A 32%-off-lane docs-only exposition edit reads as elevated loss-frequency for 🎩 operator's lane even though no code risk was actually taken; premium is restored by teaching the detector to price docsOnly commits against their true (near-zero) code-risk exposure rather than their vocabulary density.

    Open the signed attestation →

85. 2026-09-03 · 🔨 builder · offPct 18 · mistagged

    fix(digest): the nightly digest tick asks whether he is typing, not only whether we are spending

    Why it left the lane: The diff's entire vocabulary — run-digest.sh, push-digest.sh, llm-prompt.sh, quiet-gate.mjs, 'ollama run qwen2.5:7b', 'launchd', 'GPU', '4.9 GB / 100% GPU', 'session ACTIVE (quiet gate)' — is scheduling/contention/ops language with zero overlap against 🔨 builder's owned corpus (src/app, src/components, src/lib, public, build-book-html.sh, convert-chapters-to-html.mjs, voice-*-check.sh, decentralizedtalk, thetacoach-crm-mcp, vercel.json). prewalkOverlap=8 against litIntent=131 (≈6%) is the receipt's own confirmation that the reef seed built from builder's lane barely touched this reality corpus before the walk even started — the mass had nowhere in-lane to land, so it lit wide instead.

    The semantic pull: The recurring-cycle vocabulary ('the nightly digest tick', 'evicted... four times in thirteen minutes', 'the tree grew back', 'a scheduled job is still calling through the gate') pulls toward a scheduling/loop-shaped Operations coordinate — plausibly the C2.Operations.Loop family in this repo's ShortLex naming (C1.Grid / C2.Loop / C3.Flow) — never toward anything in builder's UI-rendering lane. No region narrative is cached for this sha so this is inference from the diff strings against the axis names, not a read coordinate.

    Intent–reality gap: The commit's own Story claims it is 'seated in the builder room that owns scripts/digest' — but scripts/digest/**, scripts/ops/** and tests/ops/** appear nowhere in 🔨 builder's owned-surface glob. CLAUDE.md's own commit-trailer rule says Relevant-Rooms is supposed to be the 3 highest-scoring glob matches; for these two changed files no room's glob scores above zero, so the claim reads as asserted, not computed, and the receipt measured exactly that gap as offPct=18.

    • Add scripts/digest/**, scripts/ops/**, tests/ops/** to the owning room's `.workflow/rooms/*.html` 'Owned file/dir surface' glob (operator is the CLAUDE.md-implied owner of quiet-gate policy)
    • Run /recontext to recompute room ownership now that scripts/digest and scripts/ops are active churn surfaces with no current owner
    • Add a glob-overlap check to the Relevant-Rooms trailer generator: if a listed room's glob has zero intersection with the commit's changed files, print a WARN so the Story prose can't assert ownership the glob doesn't back
    • Sweep scripts/ for other directories with no owning room glob (scripts/chat/, scripts/mesh/, scripts/pmu/ are candidates) before this class recurs somewhere else
    • Audit the last 45 builder out-of-lane events for the same unowned-path signature (near-zero prewalkOverlap, ops/infra vocabulary) to bound how much of builder's drift count is this one labeling gap versus real drift

    What we learned: Room ownership is supposed to be COMPUTED (3 highest-scoring glob matches per CLAUDE.md's own commit-trailer rule), but scripts/digest/** and scripts/ops/** currently score zero against every room's glob — so when nothing scores, attribution silently falls back to Originating-Terminal (wherever the operator happened to be sitting) instead of refusing to claim a lane. An unowned directory doesn't produce an honest 'unowned' receipt; it produces a false in-lane claim for whichever terminal was open, and the drift receipt then faithfully measures that false claim as off-lane. Prior facts-only interventions in this room (offPct trending 16→46, 15→16) never touched this because none of them could see the ownership computation itself — they were diagnosing content drift when the actual defect was upstream, in which room the content was even assigned to.

    The missing guard (the deliverable): A guard on the Relevant-Rooms trailer generator (or a standalone lint over recent commits): for each room listed in Relevant-Rooms, assert its owned-surface glob has non-zero intersection with the commit's changed files; if a listed room's glob matches nothing, fail (or WARN, matching the repo's existing soft-enforcement convention for trailer format) with the specific unowned paths named, so a mistagged claim surfaces at commit time instead of being discovered later as an offPct anomaly.

    Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Right-sized in magnitude (131/119 lit cells is neither trivially shallow nor saturated against the 144-cell lattice) but walkTruncated=true flags it hit a budget cutoff before fully resolving. Concrete adjustment: re-run this commit's walk with --max-depth increased by 1-2 plies and compare prewalkOverlap and offPct — if truncation was masking a legitimate convergence toward an ops-shaped coordinate, a deeper walk should raise overlap even before the glob fix lands; if overlap stays flat, that confirms the defect is purely seed/lane, not depth. Experiments: Re-render sha 1824f7e05's panel with --max-depth +2 and diff prewalkOverlap/offPct against this run — isolates whether walkTruncated=true is contributing to the 18% or is noise · After adding scripts/digest/**, scripts/ops/** to operator's owned-surface glob, re-render this same commit against operator's reef seed and confirm offPct drops toward the 15% threshold — this is the direct falsifier for the mistagged-room hypothesis · Run the proposed glob-overlap guard as a one-off script over builder's last 45 out-of-lane events and report what fraction share this commit's signature (near-zero prewalkOverlap + zero glob intersection) — bounds how much of the room's 45-event count is labeling artifact vs real content drift

    Measured outcome (deterministic, not self-graded): room mean offPct 18→19 over 8 commits since · dictionary ablation saved 936B

    A mistagged room claim inflates 🔨 builder's measured loss-ratio for a repair it never performed while leaving the true owner's (operator's) drift signal unmeasured entirely; premium on builder's lane is only accurately priced once the glob fix routes this commit class to the room that actually did the work.

    Open the signed attestation →

86. 2026-09-03 · ☕ network · offPct 52 · mistagged

    fix(events): seal Granary card into git + correct venue — Saffron Walden CB11, not Cambridge

    Why it left the lane: The diff-stat shows only data/events/granary-summer-networking-2026-09-04.json changed (19 insertions, 1 file) — a single JSON data record, not source under docs/outreach/, scripts/outreach/, or any of ☕ network's owned globs. The receipt's own field notes say offPct measures 'share of lit reality cells outside the intent lane' — the commit message's semantic mass (git-discipline vocabulary: 'AXIOM 1', 'witness corollary', 'sealed_at', 'control route', 'CLAUDE.md ONE DOOR', 'event-plan.mjs', 'reconstruct-the-average') pulls hard toward B3.Tactics.Signal / C1.Operations.Grid (governance/verification prose), while the reality-lane payload is a single-file JSON edit with zero matching structural surface in the room's declared owned-surface list. litIntent=127 vs litReality=84 with only prewalkOverlap=3 confirms the message text lit far more of the lattice than the actual changed bytes could ground.

    The semantic pull: Strings actually in the commit message that drag placement toward B3.Tactics.Signal/C1.Operations.Grid: 'control route', 'AXIOM 1', 'witness corollary', 'sealed_at', 'gitignored', 'event-plan.mjs', 'reconstruct-the-average', 'regression guard', 'bf-5002' — these are governance/pipeline-integrity terms belonging to the CLAUDE.md rules layer, not to a venue-correction data edit. The reds cluster on B1/B2/C1/C2/C3 (Speed×Grid/Loop/Flow, Deal×Grid/Loop/Flow) — process/topology/exchange language — which is exactly what a message full of axiom-citation and audit-trail prose lights up, independent of the one-line venue string fix it actually performed.

    Intent–reality gap: The commit message claims a careful two-part audit-and-fix (venue correction + AXIOM-1 seal-receipt discipline) with extensive self-justifying prose citing CLAUDE.md rules verbatim; the receipt measured a single-file, single-field JSON edit with almost no code-surface breadth (litReality=84, walkTruncated=true) — the message's argumentative mass vastly outweighs the tiny, room-invisible reality it describes.

    • Add a `data/events/*.json` glob to ☕ network's (or whichever room owns event cards) declared owned-surface in `.workflow/rooms/*.html` so a future venue-correction commit has a home lane to land in.
    • Move the AXIOM-1/witness-corollary rationale out of the commit body into the `Story:` trailer per COMMIT MESSAGE STRUCTURE, keeping the subject+body proportional to the diff size.
    • Add a `url`/`source_url` field to the event-card JSON schema so a future fetch-and-diff guard (noted as a real gap in this commit) becomes possible without a network call in CI — store the last-verified full_address string alongside venue for drift comparison.
    • Confirm which room actually owns `data/events/` — right now no room's globs list it, so every event-card commit will mistag by construction until one claims it.

    What we learned: A commit whose diff is a single small data field can still drag the drift receipt's lit-mass into an unrelated governance lane purely through commit-message vocabulary density — the receipt is measuring the MESSAGE's semantic gravity as much as the code's, and a message that quotes CLAUDE.md axioms at length will out-lit a one-line JSON edit every time, regardless of which room is 'correct.' This is distinct from the three prior lessons for this room (all facts-only, no mechanism captured) — this is the first sensemaking pass to actually name commit-message-verbosity-vs-diff-size as the pull mechanism.

    The missing guard (the deliverable): A pre-commit or receipt-time check comparing commit-message length/rule-citation-density against diff line-count for `data/**/*.json`-only commits — if message body exceeds e.g. 8x the diff's line count and cites ≥2 CLAUDE.md rule names, require a `Story:` trailer split so the governance argument doesn't compete with the data-fix for lattice placement.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. walkTruncated=true with litReality=84 on a single-file 19-line diff suggests the walk is oversaturating relative to the actual code surface — it's spending its ply budget spreading across governance-adjacent cells that the message vocabulary seeded rather than terminating early on the genuinely tiny reality surface. Concrete adjustment: for docsOnly/single-small-file diffs, cap walk plies at 1 (not the current 2) when the reality file count is 1 and file size is under a threshold, so the walk can't outrun the actual changed-code mass. Experiments: Run the same commit through the pipeline with the commit-message body truncated to the first 3 lines (drop the AXIOM/rule-citation paragraphs) and compare offPct — if it drops materially, that's direct evidence message-verbosity (not code) drives placement here, confirming the reef_signal_pct estimate above. · A/B the walkTruncated=true cap: re-run with --max-depth 1 vs the current setting on this sha and measure whether litReality shrinks toward the diff's true 1-file footprint without losing legitimate lane hits. · Measure prewalkOverlap across the last 20 docsOnly=true commits and correlate with offPct — if low prewalkOverlap consistently predicts high offPct, that's a cheap pre-walk early-warning signal worth surfacing in the receipt itself.

    Measured outcome (deterministic, not self-graded): room mean offPct —→9 over 7 commits since · dictionary ablation saved 855B

    An unbounded-argumentation commit message on a single-field data fix inflates this lane's measured breach rate without any actual cross-lane work having occurred, which is a false-positive loss event — premium is restored by shipping the message-proportionality convention and the owned-surface glob so the same class of commit stops registering as a claim.

    Open the signed attestation →

87. 2026-09-03 · 🔒 vault · offPct 25 · mistagged

    feat(notify): batched visit digest — one email per run, by locale then IP, funnel painted twice

    Why it left the lane: data/rooms.json's vault entry carries the literal bare token "IP" inside vector_keywords (alongside patent, examiner, claims, filing) — vault's whole reason for being is intellectual-property/patent work. The diff is saturated with the OTHER meaning of the same two letters: `ip_address` as a fetched column, the `byIp` map, the table titled 'By IP', the line 'distinct IPs', and the docblock's own words 'batching by locale then IP'. Nothing in the diff touches patents, claims, or filings — the string collision is bare-token vs bare-token, no bigram context ("IP address" vs "IP filing") to disambiguate them. That collision is the semantic pull dragging ~25% of the reality mass into vault's coordinate (A1 Strategy.Law) alongside the correctly-lit builder/operator cells that scripts/notify/*.mjs actually occupies.

    The semantic pull: The bare token "IP" is present in vault's vector_keywords with no requirement that it co-occur with patent/claim/filing language, so any diff using "IP" to mean Internet-Protocol address scores against the same keyword as a diff about intellectual-property filings — a homonym collision, not a topical overlap.

    Intent–reality gap: The commit claims a notify/analytics feature — a batched visit digest with locale and IP tables, filed entirely under scripts/notify/ and tests/notify/, none of which touches vault's owned surface (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/) — yet the receipt measured 25% of the lit reality mass landing in vault's lane, and the Relevant-Rooms trailer lists vault first with zero glob overlap on the changed files.

    • data/rooms.json: replace vault's bare "IP" vector_keyword with "IP filing" / "intellectual property" (see spec_rewrite)
    • Audit every room's vector_keywords in data/rooms.json for other bare 2-4 letter tokens that double as common technical abbreviations (e.g. check builder/operator/architect lists for similar collisions) — this incident is a class, not a one-off
    • Whatever consumes vector_keywords for commit Relevant-Rooms suggestion should require the token to be part of a domain bigram ("IP filing", "IP claim") when the bare form is a known homonym, mirroring the AXIOM 0 pmu-rust/PMU homonym-naming pattern already established in CLAUDE.md
    • Re-run the Relevant-Rooms computation for 1b12d8ca5 with the corrected keyword list and confirm vault drops out of the top-3 (builder/operator should surface instead)
    • Add a homonym stop-list (IP, PMU, key, vault, and any other token that already collides in this repo) referenced by whichever script scores room-content overlap

    What we learned: e94a8cd14 diagnosed 'the trailer generator falls back to Originating-Terminal even with zero glob match' and measured a real improvement, but this event shows a SECOND, independent mechanism producing the identical symptom (vault mistagged with zero owned-surface overlap): a bare homonym token sitting inside a room's own keyword list in data/rooms.json. Fixing the trailer-generator's fallback logic does not touch this — the reef can independently re-derive a vault match from content alone, via a keyword collision, even if the fallback bug were fully patched. Two different root causes converge on the same visible symptom, which is why the prior fix's measured improvement (21→12) didn't hold — this room's 24 out-of-lane events in 14 days include cases the previous lesson's fix cannot reach.

    The missing guard (the deliverable): A test asserting no room's vector_keywords array in data/rooms.json contains a bare token (length ≤4, all-caps or common-word) that is also a load-bearing field name or common abbreviation elsewhere in the repo (ip_address, key, vault, etc.) without requiring bigram context — run it against the historical false-positive (this commit, reintroducing bare "IP") and confirm it goes red, then green once the keyword is disambiguated.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not saturated — litReality (137) is well under the 144-cell ceiling and walksReality/litReality both equal 137 (no truncation loss despite walkTruncated=true), so the walk completed a full, non-clipped traversal. The adjustment isn't depth, it's input: gate the reef seed for a homonym-flagged keyword (see missing_guard) BEFORE the walk starts, so the walk propagates a corrected seed instead of faithfully amplifying a mistagged one. Experiments: Re-run this commit's reef seeding with vault's "IP" keyword replaced by "IP filing"/"intellectual property" and measure whether offPct drops below the 15 threshold — a direct ablation on the actual defect · Build a homonym stop-list (IP, PMU, key, vault + any token flagged by the new guard) and re-seed the last 30 commits' reef placements, counting how many previously-mistagged Relevant-Rooms suggestions change · Add a bigram-context requirement to keyword scoring generally: a bare token under 5 characters only counts toward a room's score if it co-occurs within N tokens of a second room-specific keyword, then measure false-positive rate on the last 90 days of commits before/after

    Measured outcome (deterministic, not self-graded): room mean offPct —→25 over 4 commits since · dictionary ablation saved 890B

    A room with 24 out-of-lane events in 14 days and a keyword collision that will keep firing on any networking/analytics commit reads as an unreliable attribution channel, which raises this lane's loss-ratio on every future receipt until the bare-"IP" token is disambiguated and the homonym guard ships.

    Open the signed attestation →

88. 2026-09-03 · 🎭 performer · offPct 41 · mistagged

    routes(exposition): clean routes — /exposition/:slug rewrites to the static html

    Why it left the lane: The diff is a next.config.js rewrite-rule edit plus three static-HTML edits in public/exposition/ — routing/infra terms ('rewrites', 'source', 'destination', 'slug pattern') and content-nav terms pull the walk into the C1.Operations.Grid / topology-and-routing region and the B-row Tactics blocks (Speed/Deal/Signal × Grid/Loop/Flow) rather than staying inside the git-discipline B3.Tactics.Signal × C.Operations lane the commit message frames itself in. The message talks like a commit-hygiene note ('one generic rewrite... following the file's existing air-receipt pattern') but the actual payload is a routing-topology change, which is a different semantic neighborhood than 'git discipline.'

    The semantic pull: Strings 'rewrites()', 'source:', 'destination:', '/exposition/:slug([a-z0-9-]+)' in next.config.js are routing/topology vocabulary that maps to Operations.Grid (C1) and the Tactics.Speed/Deal (B1/B2) × Operations (C1/C2/C3) blocks the receipt encircled; none of that vocabulary lives in B3.Tactics.Signal (the git-discipline/commit-signal lane this receipt is nominally scored against), so the reality walk lit 28 red cells outside the claimed pixel while the commit message never mentions routing, rewrites, or URL topology at all.

    Intent–reality gap: The commit message frames this as a small, self-similar hygiene change ('one generic rewrite... following the existing pattern') scoped like a git-discipline/process commit, but the receipt measures the actual payload as a routing-topology change (a Next.js rewrite rule + three nav-link edits) that never touches git discipline at all — the claim and the measured content are talking about two different things.

    • Add an owned-surface glob for next.config.js rewrites + public/exposition/*.html to .workflow/rooms/architect.html (or builder.html) so routing changes stop defaulting to whichever terminal typed them
    • Update the commit trailer convention note in CLAUDE.md's COMMIT MESSAGE STRUCTURE section to require Relevant-Rooms reflect the touched surface (next.config.js → architect/builder), not the room narrative in Story
    • Cross-check the two other exposition pages' commits (actuary-variables, waterfall-knee, the-third-address-space) for the same mistagging pattern before they compound the room's offPct trend
    • Add a routing-rules test asserting /exposition/:slug regex excludes dots (so .html literals still resolve) as a regression guard, since the diff comment claims this but nothing enforces it

    What we learned: Room attribution failures aren't only about missing globs or ignored trailers (the two prior lessons) — here the OWNERSHIP CLAIM ITSELF, freely chosen by the author in Relevant-Rooms, was wrong: a routing/infra change was voluntarily framed and trailer-tagged as a git-discipline/process commit because the author reasoned by analogy to 'the file's existing pattern' rather than by what surface the diff touches. The detector has no way to catch a self-authored trailer that misclassifies its own content.

    The missing guard (the deliverable): A pre-commit check that diffs the changed file paths against .workflow/rooms/*.html owned-surface globs and WARNS when the Relevant-Rooms trailer doesn't include the room(s) that actually own the touched paths (next.config.js + public/exposition/* → architect/builder, not performer) — catches self-mistagging before the receipt has to.

    Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. Right-sized in raw count (121/117 lit cells, fill 0.85%, well under the 70% saturation ceiling) but MISDIRECTED — both sides walked a normal-sized area, they just walked disjoint areas. Concrete adjustment: when prewalkOverlap=1 but offPct is still high, that's a signal the intent seed itself (drawn from commit message + trailers) was too narrow/self-referential; widen the intent-side seed to include the diff's own changed-file paths (not just the message text) so intent and reality walk from a shared anchor set. Experiments: Seed the intent-side walk from BOTH the commit message AND the list of changed file paths (weighted lower), then re-run this sha and check whether offPct drops — tests whether path-aware seeding closes self-authored-message gaps · Run the pre-commit owned-surface-glob check (proposed in missing_guard) in dry-run over the last 90 days of commits and count how many would have flagged a Relevant-Rooms mismatch before this one did — measures the size of the backlog this class represents · Ablate prewalkOverlap as a pre-filter: when prewalkOverlap=1 AND offPct>threshold, tag the event 'mistagged' automatically before sensemaking runs, and check over the next 20 flagged events whether that auto-tag matches the human/LLM boundary_failure call ≥80% of the time

    Measured outcome (deterministic, not self-graded): room mean offPct 15→43 over 3 commits since · dictionary ablation saved 762B

    An out-of-lane routing change wearing a git-discipline trailer raises this lane's loss ratio by making the room-attribution signal noisy (self-tagged claims the detector can't independently verify); premium is restored by moving the ownership claim from free-text Story/Relevant-Rooms prose to a path-derived, pre-commit-checked glob match.

    Open the signed attestation →

89. 2026-09-03 · 📐 architect · offPct 18 · structural

    feat(raise-deck): the three instruments get their own row — trigger and buyer each

    Why it left the lane: The diff's real vocabulary is a mashup of two domains: insurance/finance nouns ('affirmative policies', 'parametric bonds', 'lane-adherence derivatives', 'underwriter', 'ratable underlying', 'boundary-crossing count', 'strike', 'premium') and Tailwind/CSS layout tokens repeated across nearly every added line ('text-[clamp(0.72rem,1.2vw,0.9rem)]', 'border-t-2 border-[#3a352c] pt-4', 'mt-2 text-[clamp(0.92rem,1.45vw,1.12rem)]'). Neither vocabulary class semantically matches architect's declared globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md) — those are prose-strategy docs, not React/Tailwind source. prewalkOverlap=3 confirms the reef seed barely touched intent before the walk even started; the walk then expanded reality to full saturation (litReality=144/144, walkTruncated=true) from that thin seed, dragging placement into cells no architect doc anchors.

    The semantic pull: Cross-domain diction ('Bought by...', 'strike', 'threshold', repeated Tailwind class fragments) plus a fully-saturated, truncated walk pulled reality mass across the whole 144-lattice instead of settling near architect's narrow docs/strategy footprint (litIntent=127 vs litReality=144).

    Intent–reality gap: The commit message frames this as architect's strategy-content repair (fixing the ratable underlying, structuring the instrument band) and the trailers name architect as Originating-Terminal — but every byte actually changed lives in src/app/deck/raise/RaiseDeckClient.tsx and tests/deck/raise-deck.test.mjs, files outside all of architect's, operator's, and vault's declared owned-surface globs, so the receipt measured 18% of lit reality mass with nowhere in the intent fence to land.

    • Add src/app/deck/raise/** and tests/deck/** to architect's Owned file/dir surface in .workflow/rooms/architect.html
    • Extract instrument copy strings from RaiseDeckClient.tsx into docs/strategy/raise-deck-instrument-copy.md, imported by the component
    • Ship the room-drift-<room>-check.sh fix three rooms already independently spec'd per ff09774e5, reading owned-surface globs (not just Originating-Terminal) against changed paths
    • Add a coverage-audit test that lists every file touched by the last 20 architect-tagged commits and asserts each falls under a Relevant-Room glob, flagging any orphaned path before it accrues more out-of-lane events
    • Note in CLAUDE.md's room-ownership section that deck TSX pages are a shared architect(content)/builder(layout) surface until the content extraction above ships

    What we learned: ff09774e5 fixed reading Relevant-Rooms instead of Originating-Terminal alone (20→14 offPct), but this event shows that fix is insufficient when the true landing surface is orphaned by EVERY named room: architect, operator, and vault are all named in this commit's trailers, and none of their owned-surface globs cover src/app/deck/raise/** or tests/deck/**. The raise deck has been edited every round (r1 through r10) as architect's core work, yet no room glob has ever claimed its actual code path — a structural gap independent of which trailer the detector reads.

    The missing guard (the deliverable): A coverage-audit test (e.g. tests/ops/owned-surface-covers-active-code.test.mjs) that, for the last N commits carrying a given Originating-Terminal, asserts every changed path falls inside that room's OR a Relevant-Room's declared owned-surface glob — failing today specifically for architect + src/app/deck/raise/** + tests/deck/**, forcing the glob to be extended before the next round ships.

    Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Over-saturated: litReality=144/144 is a ceiling artifact (every lattice cell lit), consistent with the AR-ledger's 'unguided flood' anti-pattern, and walkTruncated=true confirms it never settled on its own. Adjustment: lower --max-depth for this ingest class, or pre-filter the reality corpus of generic Tailwind utility-class fragments before walking, so the walk can terminate before hitting full saturation and offPct reflects real semantic spread rather than a truncation artifact. Experiments: Re-run this same sha's reality ingest with --max-depth reduced by 2 plies; if litReality still saturates at 144, the corpus (not depth) is the culprit · Strip repeated Tailwind utility-class strings (grid gap-, md:grid-cols-, border-t-2, text-[clamp() from the reality corpus before ingest and re-render this sha's panel; measure how much of the 18% offPct is CSS noise vs real insurance-vocabulary drift · Compute prewalkOverlap across all 173 of this room's 14-day out-of-lane events and check whether it clusters low specifically on .tsx/.test.mjs file commits — if so, TSX file types systematically under-seed architect's reef overlap and the fix generalizes past this one commit

    Measured outcome (deterministic, not self-graded): room mean offPct 10→7 over 4 commits since · dictionary ablation saved 1255B

    173 out-of-lane events in 14 days on this room's highest-frequency commit path inflate architect's measured loss ratio as a file-surface artifact rather than real strategic drift; extending the owned-surface glob (and eventually extracting deck copy into docs/strategy/) restores the lane so future frequency counts reflect genuine boundary crossings, not an orphaned code path.

    Open the signed attestation →

90. 2026-09-03 · 🎭 Alacritty Performer · offPct 26 · mistagged

    post(blog): envelope post factual hardening — the incident narrated as Ismail's telling, press-safe verbs, undo-impossibility added

    Why it left the lane: No region narrative was cached for this sha, so the coordinates the mass actually landed in aren't nameable — but the lane mismatch is legible from the surface list alone: the diff touches exactly one file, src/content/blog/2026-09-03-the-envelope-needs-a-countable-edge.mdx, editing prose like 'as Ismail tells it, the agent works its way past the booking window' and 'slipped the booking window' and adding 'The underlying incident is real and was covered widely in August 2026 ... could not undo the stranger's cancellation once made.' None of that vocabulary — gym-class incident, Ismail, booking window, press coverage, undo-impossibility — appears anywhere in Performer's owned-surface globs (docs/decks/, docs/strategy/deck/, fundraising, munich-re, event-, demo-ops). prewalkOverlap=11 against litIntent=134 (~8%) confirms the intent-side seed pool barely touched this room's reef before the walk even ran, which is consistent with the content simply not belonging to any pitch/deck/fundraising vocabulary — the walk then had almost nothing on-lane to anchor to and its litReality=90 spread landed 26% outside whatever thin lane it found.

    The semantic pull: The pull isn't toward a specific wrong coordinate so much as away from Performer's reef entirely — 'Ismail's telling,' 'press-safe verbs,' 'AXIOM 1 walking,' 'undo-impossibility' are blog-editorial/factual-hardening register, structurally disjoint from 'deck,' 'fundraising,' 'munich-re,' 'event-,' 'demo-' terms that seed this room's lane.

    Intent–reality gap: The commit message frames this as legitimate work ('the placement lane identified the underlying incident... no vendor named, per the register rule') and only carries an Originating-Terminal trailer (🎭 Alacritty Performer) with no Relevant-Rooms line — but the receipt measured the diff against Performer's owned-surface reef anyway, and the reality is a pure blog-content edit that shares zero owned-surface globs with that room.

    • Audit Performer's other 3 out-of-lane events from the last 14 days for the same missing-Relevant-Rooms / content-commit pattern — if most are blog/content commits authored from Alacritty, that confirms this is systemic rather than a one-off for 38ee117dd
    • Decide and document in data/rooms.json (or .workflow/rooms/*.html) whether any of the 9 canonical rooms actually owns src/content/blog/** — if none does, codify 'unscoped' as a legitimate, non-penalized lane state
    • Add a warn-only pre-commit check: a commit touching src/content/blog/**.mdx with no Relevant-Rooms trailer prints a reminder, mirroring the existing Six-Needs/lexical warn pattern
    • Update the lane-attribution code path (wherever offPct's room assignment happens) to prefer Relevant-Rooms / glob-match over bare Originating-Terminal
    • Add the missing Persona-Intent trailer requirement check for content commits (currently unenforced per 'Format violation does NOT block commit')
    • Ship the guard test named in missing_guard below in the same commit as the code_rewrite

    What we learned: Originating-Terminal (physical room at commit time) is being used, in practice, as the drift-lane a commit is scored against — even though CLAUDE.md explicitly documents that Originating-Terminal and content-relevance are allowed to diverge ('that's fine and worth recording'). Any blog-content commit typed from Alacritty will therefore ALWAYS max out-of-lane against Performer's deck/fundraising reef, regardless of how well-written or well-scoped the content itself is — this is the structural crack the three prior facts-only passes (offPct trending 30→48→30, never improving) never named because none of them had the room's owned-surface list to compare against.

    The missing guard (the deliverable): A test asserting the receipt's lane-attribution step selects its comparison reef from Relevant-Rooms (or an owned-surface glob match) rather than from Originating-Terminal alone — fed a synthetic commit with Originating-Terminal=Performer, a diff touching only src/content/blog/**, and no Relevant-Rooms trailer, it must assert the commit is scored as unscoped/no-lane rather than against Performer's reef.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true with litIntent=134 vs litReality=90 (67% fill) suggests the walk hit its ply/time budget before fully converging — right-sized-but-clipped rather than saturated or too shallow. Concrete adjustment: re-run this sha's walk with --max-depth raised one step and diff the resulting offPct against 26 to see whether truncation itself is inflating the drift reading or whether it converges to a similar figure. Experiments: Re-score sha 38ee117dd against an 'unscoped/no owned-surface match' pseudo-lane instead of Performer's reef and confirm offPct collapses toward the docsOnly/direction-only floor — this directly tests the reef-mismatch hypothesis · Pull prewalkOverlap + offPct for every Performer commit in the last 14 days, split by whether the diff matches any of Performer's owned-surface globs; if non-matching commits systematically show low prewalkOverlap and high offPct, that's the receipt for the structural mismatch, computable from existing ndjson without a new run · Re-run this sha with --max-depth raised past the current truncation point and compare offPct to the shipped 26 to isolate how much of the reading is walk-truncation artifact vs genuine reef mismatch

    Measured outcome (deterministic, not self-graded): room mean offPct —→9 over 4 commits since · dictionary ablation saved 585B

    A repeat, unfixed pattern of a room's reef absorbing content it structurally cannot own inflates Performer's measured loss-ratio with false-positive drift, which either overprices this lane's premium or trains the underlying model to ignore its own offPct alarms — both restored by scoping the lane read to Relevant-Rooms/glob-match so the number reflects actual cross-lane work, not terminal-of-origin noise.

    Open the signed attestation →

91. 2026-09-03 · 🔒 vault · offPct 22

    fix(deck/raise): Peter needs Paul — restore the pair, because the pair is the mechanism

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→40 over 4 commits since

    Open the signed attestation →

92. 2026-09-03 · 🎭 performer · offPct 46

    routes(exposition): single-word doors — /substrate /knee /variables

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→43 over 3 commits since

    Open the signed attestation →

93. 2026-09-03 · 🎩 operator · offPct 24 · structural

    fix(iamfim): one leading cell before debatable? so its letters sit over insurable's

    Why it left the lane: The new comment block inserted at page.tsx:196-203 contains the literal token "operator:" followed by a quoted dictation ("make a white space ... so the letters line up") plus a date stamp (2026-09-03) and a restated decision history ("Putting the ¿ back is one edit here"). Per the repo's own PMU ingest rule (comments + identifier-words feed the REALITY corpus, semantic not raw), that text is walked as reality content. The literal string "operator" is also the exact name of this lane's own room, and the surrounding prose reads like a decision-log/spec entry rather than implementation code — both properties pull reef mass toward doc/room-name coordinates that sit outside the narrow 'iamfim-landing CSS glyph fix' lane the commit message actually claims.

    The semantic pull: Two concrete strings from the diff: (1) the citation "operator: \"make a white space ... so the letters line up\"" — token-collides with the room name itself; (2) the dated historical narration "(2026-09-03, ... ) ... Putting the ¿ back is one edit here" — shaped like a decisions-ledger entry, not code. Both sit inside a JSX comment in src/app/iamfim-landing/page.tsx, a file with zero relationship to the operator room's owned surface (data/trust-debt/, docs/01-business/pipeline/, docs/01-business/accounting/, public/dashboard*).

    Intent–reality gap: The commit claims a single-cell CSS/typography alignment fix scoped narrowly to two <p> tags in one landing-page file. The receipt measured 24% of lit reality mass landing outside that lane — driven not by the actual code change (one &nbsp;) but by a multi-line narrative comment that reads and reef-matches like intent/spec prose quoting the room's own name.

    • Trim the page.tsx comment per code_rewrite (drop date + operator-quote + restated history).
    • Add the CODE COMMENT SCOPE rule to CLAUDE.md with the guard named below.
    • grep the repo for other `<room-name>: "..."` citation-style comments (operator/architect/builder/performer/navigator/voice/vault/laboratory/network) that may be silently mis-attributing rooms the same way.
    • Add a strip step in the reality-ingest module (wherever comments are SimHash'd for the walk) that removes name-citation + date spans before reef-seeding.
    • Re-render this commit's panel after the strip and confirm prewalkOverlap moves off 0 and offPct drops toward the room's 15 threshold.

    What we learned: PMU's own ingest rule — comments count as REALITY-semantic content — means a source comment that quotes a person and cites a date behaves like a spec/decision-log entry and can literally token-collide with a room's own name, pulling reef placement toward that room independent of the actual file touched. Prior lessons for this room were facts-only with a worsening trend (10→24, 21→26); this is the first time the mechanism has been traced to a literal name-citation token inside a code comment rather than a genuine cross-lane file touch.

    The missing guard (the deliverable): A guard that scans staged diffs for JSX/TSX comment additions matching a name-citation pattern (`\b(operator|architect|builder|performer|navigator|voice|vault|laboratory|network)\s*:\s*["“]`) and either strips that span from the reality-ingest corpus before reef-walking, or fails a lint step asking the author to move it to the commit message — e.g. tests/pmu-simulator/no-name-citation-in-comments.test.mjs guarding the strip in the ingest script.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent (105) and litReality (109) are close in magnitude, arguing the walk is behaving symmetrically and not saturated — but walkTruncated=true on a one-file, one-glyph commit suggests the max-depth budget (presumably tuned for pipeline-scale commits) is clipping small UI commits before the walk can show whether the out-of-lane cells are stray leaves or persistent anchors. Adjustment: re-run this commit with --max-depth incremented by one ply and check whether offPct 24% collapses toward the 15 threshold (stray leaf) or holds (real anchor). Experiments: Strip the citation span (/\b\w+:\s*["“][^"”]*["”]/) from the comment text, re-run this commit's reef seeding, and check whether prewalkOverlap moves off 0 — confirms or kills the citation-token seed-divergence hypothesis. · Re-run the ballistic walk on this exact commit with --max-depth+1 (walkTruncated is currently true) and record whether offPct converges toward 15 or stays ≥24 — decides truncation-artifact vs. real placement. · Ablation-test down-weighting JSX/TSX comment blocks over N words relative to code identifiers in the reality SimHash, then re-score this room's last 14 out-of-lane events to see whether mean offPct drops — tests whether prose-heavy comments are the repeat driver behind this room's rising trend (10→24, 21→26).

    Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 6 commits since · dictionary ablation saved 763B

    This single event is cosmetic (low sigmaDrift, low driftPct) and barely dents the operator lane's loss ratio on its own, but the room's rising 14-day trend (10→24, 21→26 across prior interventions) means premium keeps climbing on a false signal — a citation-token collision, not real scope creep — until the ingest-strip guard ships and lets the room's offPct track genuine drift instead of comment-text noise.

    Open the signed attestation →

94. 2026-09-03 · 🎤 voice · offPct 29 · mistagged

    feat(bookclub): backstage funnel measured end-to-end — interior lanes + the read door

    Why it left the lane: Every changed path in this diff sits on ground explicitly deeded to other rooms in the .workflow/rooms/*.html registry itself: scripts/bookclub/backstage-funnel.mjs falls under `scripts/bookclub/`, which messages-network.html states was 'Claimed 2026-08-24 (operator: book club is a strategy similar to market contact)... It sat unowned while its ledgers accumulated' — that is ☕ Network's line, verbatim, not voice's. src/components/iamfim/BackstageSignup.tsx sits under `src/components/`, which terminal-voice.html's own Forbidden surface names outright: 'src/app/, src/components/ at the infrastructure level — owned by 🔨 Builder.' tests/ops/backstage-funnel.test.mjs is generic hook/test infra, again Builder's grid. The vocabulary that drove the walk — 'mirror-sync.mjs', 'user_activity_logs', 'subscribers.metadata', 'CUTOFF', 'data-lane="bookclub-submit"' — is Operations/measurement lattice language, nowhere near voice's book/blog anchors (books/tesseract/chapters, src/content/blog). Zero of the 201 inserted lines touch voice's owned globs.

    The semantic pull: scripts/bookclub/, src/components/, tests/ops/ are three separate ownership claims (Network, Builder, Builder) that all pull away from voice's books/tesseract + src/content/blog anchors; none of the diff's real strings (backstage-funnel, mirror-sync, SessionShapeRecorder, CUTOFF) appear anywhere in voice's owned-surface vocabulary.

    Intent–reality gap: The commit message claims 'the third beat of the backstage-pass session' as ops/measurement work and even tags Relevant-Rooms as 🔨 builder, ☕ network, 🧪 laboratory — never voice — yet Originating-Terminal is stamped 🎤 voice, and the receipt scores offPct against the terminal room's owned surface, so a commit that CLAUDE.md's own commit-structure rule says is fine to diverge on ('a commit's originating terminal and its file-relevance to rooms can diverge — that's fine and worth recording') is instead measured as a 29%-over-threshold violation.

    • Add the spec_rewrite sentence to CLAUDE.md's COMMIT MESSAGE STRUCTURE section (same commit as the pipeline fix, per CODE DEFINITION-OF-DONE)
    • Locate the receipt/room-ledger script that attributes offPct to a room from commit trailers (likely under scripts/pmu/ or a room-drift aggregator) and patch it to parse Relevant-Rooms first, falling back to Originating-Terminal only when Relevant-Rooms is absent
    • Re-score the prior 90 out-of-lane 🎤 voice events from the last 14 days against Relevant-Rooms to measure what fraction are this same terminal-vs-ownership mistag class — that number bounds the expected fix impact
    • Confirm scripts/bookclub/** is listed under messages-network.html's owned surface (it already is, per the 2026-08-24 claim) and src/components/** under iterm2-builder.html (it already is) — no glob work needed, only the attribution fix
    • Add the missing_guard test below in the same commit as the pipeline patch

    What we learned: The per-room drift receipt currently uses Originating-Terminal (physical typing location) as the lane key, but CLAUDE.md's own commit-structure rule explicitly licenses that trailer to diverge from file-relevance — meaning any legitimate cross-room commit typed from a given terminal is STRUCTURALLY GUARANTEED to register as out-of-lane for that room, independent of code quality. This is the first sensemaking pass recorded for 🎤 voice (all three priors were facts-only), and it reframes the room's 90-event/14-day count: some unknown fraction may be this mechanical mistag rather than genuine content-lane drift, which the room's two prior facts-only interventions (offPct 27→39, 0→36) could not have caught since they had no diagnosis to work from.

    The missing guard (the deliverable): A test that renders a synthetic commit with Originating-Terminal=roomA and Relevant-Rooms=roomB (roomB's owned globs matched, roomA's not) and asserts the receipt's lane-threshold check attributes offPct against roomB, never roomA — proving the room used for measurement is always the file-ownership room when the two trailers disagree.

    Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. litIntent/walksIntent=96/96 and litReality/walksReality=132/132 — every walked step lit, a fully saturated walk with zero wasted branches on both sides — but walkTruncated=true means it was cut off by a time/depth budget while still actively lighting new cells, not because it converged. Concrete adjustment: raise --max-depth (or the time budget) specifically for reality-side ingestion on infra/ops-heavy commits like this one, since reality already outgrew intent by 38% before truncation and the true out-of-lane spread past that point is unmeasured. Experiments: Re-run this commit's reef seed using a corpus built from Relevant-Rooms (builder+network+laboratory) instead of Originating-Terminal (voice) and measure whether prewalkOverlap rises materially above 6 — a rise would directly confirm the mistag hypothesis is reef-visible, not just a lane-threshold artifact · Split the room's last 90 out-of-lane events by whether the room appears in Relevant-Rooms or not, and measure what fraction of the 90 falls in the Relevant-Rooms-absent bucket — that fraction is the ceiling on how much this single attribution fix can reduce the room's offPct trend · Re-render this commit's triptych scored against 🔨 builder's owned surface instead of 🎤 voice's and measure whether offPct drops below the 15 threshold — expected yes, since scripts/bookclub/ and src/components/ are named in builder/network's owned surface already

    Measured outcome (deterministic, not self-graded): room mean offPct 15→30 over 11 commits since · dictionary ablation saved 1024B

    A mistagged-lane event like this inflates 🎤 voice's apparent loss-ratio without any underlying authorship defect — the work is correctly placed in Network's and Builder's own claimed territory — so the room's premium is being penalized for an attribution bug, not for drift; restoring Relevant-Rooms-keyed lane scoring removes that false-positive drag and lets voice's premium track only genuine content-lane divergence.

    Open the signed attestation →

95. 2026-09-03 · 🎤 voice · offPct 22

    feat(iamfim): the backstage pass — book-club signup CTA below the two-door fork

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 15→0 over 4 commits since

    Open the signed attestation →

96. 2026-09-03 · 🔒 vault · offPct 40 · mistagged

    feat(sticker): one campaign sticker, one place — and it is the design that was chosen

    Why it left the lane: Every changed path in the diff — src/components/CampaignSticker.tsx, src/app/deck/dentry/DentryDeckClient.tsx, src/app/deck/raise/RaiseDeckClient.tsx, src/app/iamfim-landing/page.tsx, tests/components/campaign-sticker.test.mjs — is component-extraction/marketing-surface work. The commit message itself is saturated with 'sticker', 'campaign', 'landing page', 'deck', 'artwork', 'QR' — none of which are patent/legal vocabulary. The reality walk lit 127 cells clustered around that UI/product-surface territory while the claimed lane (vault: docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/) matched zero of the five changed files, so 40% of the lit reality mass fell structurally outside the intent lane's glob footprint before any semantic judgment was even possible.

    The semantic pull: The pull terms are the diff's own repeated nouns: 'CampaignSticker', 'iamfim-landing', 'DentryDeckClient/RaiseDeckClient', 'the disc', 'the sentence' — a cluster of marketing/UI identifiers with no legal or patent cognate anywhere in the five touched files, so the walk had nothing vault-shaped to grip onto even as a false positive.

    Intent–reality gap: The commit's own trailers claim Originating-Terminal: 🔒 vault and list vault first in Relevant-Rooms, but the receipt shows the actual work (sticker component extraction, two decks, a landing page) has zero overlap with vault's owned surface — the claim is about which terminal was open, not about what was built, and the pipeline reads the former as if it were the latter.

    • Locate every call site that emits the Relevant-Rooms trailer (hook + any interactive commit path) — e94a8cd14's fix may have patched only one of several
    • Add tests/rooms/relevant-rooms-glob-match.test.mjs asserting no room with a zero owned-surface-glob score against the commit's changed files can appear in Relevant-Rooms
    • Re-audit the 23 vault out-of-lane events in the last 14 days for the same zero-glob-score pattern to size how much of the room's current offPct trend is this one mechanism vs genuine vault-territory drift
    • Add an explicit rule to CLAUDE.md's COMMIT MESSAGE STRUCTURE section: 'Relevant-Rooms generator must exclude any room whose owned-surface glob score is zero, never fallback-include Originating-Terminal' — graduate the diagnosis into the surface every session reads, not just this ledger
    • Confirm whether e94a8cd14's fix actually shipped as code or only as a prose diagnosis in the ledger — if code, find why it regressed; if prose-only, that IS the root cause of this recurrence

    What we learned: A diagnosis recorded only in the drift ledger (e94a8cd14: 'Relevant-Rooms lists vault first despite zero glob match') is not a fix until it is code — the same mistag mechanism recurred here at HIGHER severity (offPct 40 vs 35, 23 events vs 13 in the prior window) three events later, which means the measured 21→12 improvement after e94a8cd14 was noise regressing to the mean, not correction. The pipeline's own 'MEASURED OUTCOME' field can look like progress while the underlying generator bug ships unchanged.

    The missing guard (the deliverable): tests/rooms/relevant-rooms-glob-match.test.mjs — for a representative commit, assert every room named in the Relevant-Rooms trailer has nonzero overlap between its `.workflow/rooms/*.html` owned-surface globs and the commit's actual changed files; fail if Originating-Terminal appears in Relevant-Rooms with a zero score.

    Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. Volume looks right-sized (123 vs 127 lit cells) but walkTruncated=true means the walk was cut before full convergence — with a mistagged intent lane feeding it, a truncated walk may have stopped before finding what little real vault overlap exists, inflating offPct further than a fully-converged walk would show. Rerun this sha with a higher --max-depth and compare offPct; if it drops, truncation is compounding the mistag rather than just the mistag alone. Experiments: Re-run the reef seed step gating on glob-verified room instead of the trailer's claimed room, and compare prewalkOverlap for this commit under the corrected label vs the current mistagged one · Add a pre-walk assertion: if the claimed intent-lane room's owned-surface glob score against the commit diff is zero, flag the event as MISTAGGED before the walk runs rather than after, and route it straight to the trailer-generator fix instead of a full sensemaking pass · Sample the other 22 vault out-of-lane events from the last 14 days and check what fraction share this exact zero-glob-score pattern — if it's most of them, the room's whole 'offPct trend' is measuring a generator bug, not vault's actual drift

    Measured outcome (deterministic, not self-graded): room mean offPct 37→25 over 4 commits since · dictionary ablation saved 1197B

    A room whose stated lane matches zero of its own changed files is an unpriceable signal — this lane's loss ratio can't be trusted until the trailer generator only ever claims rooms with real glob overlap, restoring offPct as a measure of actual vault drift rather than of a labeling bug.

    Open the signed attestation →

97. 2026-09-03 · 🎭 performer · offPct 50

    post(blog): twenty predictions, screened and PRICED — course G + the committed prediction ledger

    Why it left the lane: (sensemaking unavailable — facts-only intervention)

    Measured outcome (deterministic, not self-graded): room mean offPct 31→15 over 10 commits since

    Open the signed attestation →

98. 2026-09-03 · 🔒 vault · offPct 46 · mistagged

    fix(deck): Peter is introduced where he is used, and carries his own role

    Why it left the lane: The commit trailers claim vault (Originating-Terminal: 🔒 vault, Relevant-Rooms: 🔒 vault, 📐 architect, 🎩 operator) but every changed path — src/app/deck/dentry/DentryDeckClient.tsx, src/app/deck/raise/RaiseDeckClient.tsx, tests/deck/peter-is-introduced.test.mjs — matches zero of vault's owned globs (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/). Worse: vault's own room spec (.workflow/rooms/wezterm-vault.html, 'Forbidden surface') EXPLICITLY lists src/app/ and src/components/ as 'owned by 🔨 Builder'. The trailer wasn't computed from a glob match at all — it was seeded from Originating-Terminal (the physical terminal, WezTerm=vault, where the session happened to be open) and rubber-stamped into Relevant-Rooms. prewalkOverlap=2 confirms it: the reef found almost no shared vocabulary between vault's seed corpus (patent/legal prose) and this commit's actual content (deck UI copy about an AI claims assistant's authority bound), so the walk had nowhere honest to land and offPct blew out to 46.

    The semantic pull: root cause traced to scripts/post-commit-gemini-trailer.sh's hand-maintained 'hints' dictionary (~line 104), which assigns 'docs/architecture, src/app, src/components, system design' to architect — contradicting .workflow/rooms/iterm2-builder.html's canonical declaration that src/app/, src/components/, src/lib/ are Builder's owned surface. Vault's hint ('docs/patents, legal, .pem, claims') never mentions src/app either way, so nothing in the prompt path ties this commit's real files to vault — the LLM (or the authoring session mirroring the same convention) defaulted to the terminal it was physically in instead of computing overlap.

    Intent–reality gap: The commit claims this is vault's work (first-listed Relevant-Room, matching Originating-Terminal); the receipt measured 46% of the lit reality mass outside vault's intent lane because the work is 100% inside Builder's owned surface (src/app/deck/**, tests/deck/**) — a room vault's own spec explicitly forbids itself from touching.

    • Fix scripts/post-commit-gemini-trailer.sh's hints dict to derive from .workflow/rooms/*.html instead of a hand-copied table
    • Implement scripts/room-drift-vault-check.sh (spec'd at wezterm-vault.html:633, still 'not yet implemented')
    • Implement scripts/room-drift-builder-check.sh (spec'd at iterm2-builder.html:662, still 'not yet implemented')
    • Add a WARN-only glob-overlap check in hooks/post-commit on the just-written Relevant-Rooms trailer, mirroring the WIP-claims advisory pattern
    • Close self-prompts sp-454, sp-455, sp-460 — already queued in vault's own room card describing this exact defect — by shipping the guard rather than leaving them open
    • Record the corrected Relevant-Rooms for b592817d5 via git notes (never amend shared main)

    What we learned: There is no deterministic glob-check anywhere in the Relevant-Rooms production path — it is pure LLM guess, seeded by a hint dictionary (scripts/post-commit-gemini-trailer.sh) that has ITSELF drifted from the canonical .workflow/rooms/*.html owned-surface declarations (assigning src/app/src/components to architect when the canonical doc assigns them to builder and explicitly forbids them to vault). Prior lesson e94a8cd14 fixed which field downstream trusts (the READ side) but never touched how Relevant-Rooms gets produced (the WRITE side) — so the same terminal-defaults-to-relevant failure recurred, now worse (offPct 46 vs the 21→12 that was measured), because nothing upstream can check the guess against ground truth.

    The missing guard (the deliverable): tests/rooms/relevant-rooms-glob-match.test.mjs: for a commit's changed-file list and its Relevant-Rooms trailer, parse every room's real Owned file/dir surface globs out of .workflow/rooms/*.html and assert every room named in the trailer has at least one glob match against the changed files — wired WARN-only into hooks/post-commit so a mistag is visible at commit time instead of discovered days later by the drift receipt.

    Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent = 125/125 = 100% fill, litReality/walksReality = 135/135 = 100% fill — fully saturated on both sides. Given saturation, the walk correctly filled a wrong-shaped lane rather than under- or over-reaching; no depth/breadth adjustment is indicated. The fix belongs at seed selection, not walk tuning. Experiments: Over the last 30 commits carrying Originating-Terminal, compute prewalkOverlap for (Originating-Terminal-as-intent) vs (actual-glob-matched-room-as-intent) and measure whether prewalkOverlap<5 predicts high offPct — if it correlates, prewalkOverlap becomes a cheap pre-walk mistag detector that fires before the expensive walk runs · Regenerate each room's reef seed vocabulary directly from its .workflow/rooms/*.html Owned-surface + Primary-requirements-domain prose and re-measure prewalkOverlap on this same commit — if vault's seed still shows near-zero deck/UI vocabulary, that confirms the seed is correctly narrow and the trailer (not the reef) is the sole defect · Surface prewalkOverlap in the lesson-input context table for future intervention passes — it was the single strongest signal in this event and is currently computed but not shown to the diagnosing pass

    Measured outcome (deterministic, not self-graded): room mean offPct 14→34 over 6 commits since · dictionary ablation saved 1268B

    On its own this is a labeling defect, not a code defect, so it doesn't raise vault's real loss exposure — but 21 out-of-lane events in 14 days for one room erodes confidence in the room-level attribution the whole receipt pipeline prices against, and premium is restored by shipping the deterministic glob-check rather than by another one-off relabeling.

    Open the signed attestation →

99. 2026-09-03 · 🎩 operator · offPct 45 · structural

    post(clearing-house): course C names which eigenvalue the receipt races

    Why it left the lane: The diff is a single added clause inside the existing margin paragraph, and every noun-phrase in that clause is PMU-pipeline-internals vocabulary rather than pedagogy vocabulary: 'the walk is directed by the ordering', 'ShortLex index', 'directed acyclic walk', 'the discount is not what buys convergence, it buys locality', and the callout two paragraphs down cites `scripts/pmu/rc-compute.mjs` and 'ablated' / 'scattered layout'. That is the exact diction of data/pmu/pipeline/ and data/trust-debt/ — the room's own owned code surface — dropped into a blog markdown file. prewalkOverlap=4 (out of litIntent=130) shows the intent seed (drawn from the commit message's course/eigenvalue/pole framing) shares almost no vocabulary with the reality seed (drawn from the diff's mechanism prose) before the walk even starts, so the walk propagates from two nearly-disjoint starting points and the reality lobe lands outside the declared lane.

    The semantic pull: 'ShortLex index' + 'directed acyclic' + 'buys locality' + 'scripts/pmu/rc-compute.mjs' are pipeline-implementation strings, not course-explainer strings — they pull the reality walk toward the Operations/pipeline coordinates the same commit's own room owns in code, even though the file touched is content, not code.

    Intent–reality gap: The commit claims to be a one-sentence pedagogical clarification inside a published course (`post(clearing-house)`, Persona-Intent aimed at 'a risk actuary'), staying inside a content/explainer lane — but the receipt shows 45% of the lit reality mass sits outside that lane, because the clarifying sentence is written in the pipeline's own internal mechanism vocabulary rather than in the course's pedagogical register.

    • Add the Story-trailer acknowledgment above to this commit's message convention for future PMU-explainer posts (CLAUDE.md COMMIT MESSAGE STRUCTURE section)
    • Extract the shared 'directed acyclic / discount buys locality' correction paragraph (duplicated verbatim across this commit and 15cb70c4a9 on /substrate) into a single reusable snippet so the next correction is one edit, not two independently-drifting ones
    • Widen 🎩 operator's owned-surface glob in .workflow/rooms/operator.html to explicitly recognize PMU-architecture-explainer blog posts as part of its lane, not just data/pmu/pipeline/ code
    • Cite docs/architecture/anti-rules-ledger.md or the AXIOM 0 section by anchor link from the post itself, since that's where 'directed acyclic' and 'ShortLex ordering' are already canonically defined — this doubles as reader-facing evidence-last sourcing
    • Re-run this same paragraph's offPct against 15cb70c4a9 (the sibling correction) to confirm the drift is systematic across both posts, not commit-specific

    What we learned: A blog post that explains the PMU pipeline's own mechanism (not just references it) will always split its lit-reality mass into the pipeline's code coordinates, because the accurate explanation IS that vocabulary — this is a structural category (content-that-describes-infrastructure) the room-ownership globs never accounted for, distinct from the prior three lessons which shipped no sensemaking at all.

    The missing guard (the deliverable): tests/pmu-simulator/pmu-explainer-intent-seeding.test.mjs — seeded with df6b21727e's actual pre/post bytes (per GUARD-SEEN-TO-FAIL): asserts that when a blog diff introduces terms already defined in docs/architecture/ (ShortLex ordering, directed acyclic, criticality), the commit's intent corpus is expanded with that doc section before the walk runs, and that offPct for the fixture commit drops back under the room's 15 threshold once expanded — red on the unexpanded ingest, green once the doc-section pull is wired in.

    Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Both sides are saturated (100% lit-to-walked) and walkTruncated=true, meaning the walk hit its depth/budget ceiling rather than terminating naturally on this dense a paragraph. That's under-resourced for technical content of this density — concrete adjustment: raise --max-depth specifically for commits whose diff exceeds a term-density threshold of doc-defined mechanism vocabulary, and re-measure whether offPct changes once the walk is allowed to fully resolve rather than being cut off mid-spread. Experiments: Re-run reef seeding for this commit with the linked AXIOM-0/anti-rules-ledger section appended to the intent corpus and measure whether prewalkOverlap rises above 4 · Re-run with walkTruncated disabled (raised max-depth) on this same commit and measure whether offPct moves independent of the seeding change, to isolate truncation's contribution from seed-overlap's contribution · Diff this commit's offPct against sibling commit 15cb70c4a9 (same correction, different post) to confirm the drift is a systematic class rather than an artifact of this one file's surrounding paragraph

    Measured outcome (deterministic, not self-graded): room mean offPct —→21 over 6 commits since · dictionary ablation saved 834B

    An offPct of 45 on a one-line pedagogical fix reads as high loss-frequency on this lane's book, but because the mass lands in the room's OWN pipeline vocabulary rather than a foreign lane, the true severity is low — restoring the premium just requires the intent-seeding fix above so the receipt stops flagging legitimate architecture-explainer content as drift.

    Open the signed attestation →

100. 2026-09-03 · 🎭 performer · offPct 51 · mistagged

     fix(milestones): repair three predicates that could never flip — 8s-timeout probe, all-history -eq 0, mutable-tree read

     Why it left the lane: The diff has zero semantic file content — only data/milestones-progress.json and data/milestones.json changed, a pure data-asset commit with no code/prose to grip. The commit message's own vocabulary ('8s-timeout probe', 'append-only ndjson', 'mutable-tree read', 'git show HEAD:', 'AXIOM 1 W3/W4', 'zero-FAILED among newest-5') is infra/measurement-verb-heavy and structurally closer to Operations.Grid/Loop/Flow (build pipelines, throughput, velocity) and Tactics.Deal/Speed (exchange, timing) than to the room's actual owned surface (decks, fundraising docs, event scripts). With no diff text to anchor the walk to docs/decks/* or docs/strategy/deck/*, the gzip-NCD reef seed and the ballistic walk both drift toward the generic infra-repair vocabulary that dominates this commit's message.

     Intent–reality gap: The commit message claims a scoped, room-agnostic sprint-board repair ('repair three predicates that could never flip') with no reference to any 🎭 performer surface (decks, fundraising, event scripts) — but the receipt measured it as Originating-Terminal 🎭 performer, so the lane comparison is being made against a room this commit never claimed content ownership of in the first place.

     • Add a data/milestones*.json owned-surface glob to whichever room is the actual home for sprint-board/milestone tooling (none currently claims it — audit .workflow/rooms/*.html)
     • Add an explicit exemption class in the lane-scorer for pure-JSON/data-asset commits (diff stat shows only .json changes, zero .md/.ts/.tsx/.mdx) so they are scored against a 'data-infra' pseudo-lane rather than the terminal's content lane
     • Update scripts that compute Relevant-Rooms (room-activity.mjs) to drop a room from the trailer when that room's owned-surface globs match none of the changed paths, rather than carrying it forward by terminal habit
     • Cross-reference this event against the fdde61d21/f9e88143b lesson chain: this is the THIRD escalation of the same root cause (Relevant-Rooms union not consulted) — consider making Relevant-Rooms union-matching the intervention-fire.sh default rather than a per-event patch

     What we learned: A commit whose diff is 100% JSON data-asset (no prose, no code identifiers) has nothing for the gzip-NCD reef or the ballistic walk to grip against its true subject — the walk falls back to the commit MESSAGE's infra vocabulary, which structurally over-indexes on Operations/Tactics coordinates regardless of which room actually owns the touched files. This is a new failure mode distinct from the prior Relevant-Rooms-ignored lessons: even a perfectly-resolved Relevant-Rooms union would not save this commit, because the underlying files (data/milestones*.json) are owned by NO room's glob at all — it is the f08ec5875 'ownerless shared bookkeeping' class recurring on a new file pair.

     The missing guard (the deliverable): A test asserting every top-level data/*.json file referenced by an active script (grep scripts/ for the literal path) has at least one owning glob in some .workflow/rooms/*.html — flagging data/milestones.json and data/milestones-progress.json as currently unowned, the same class f08ec5875 already flagged for room-bifurcations.json and self-prompts.json.

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized in mechanics (128 vs 131 walks/lit is balanced, no saturation or truncation — walkTruncated=false) but starved on INPUT: a docsOnly=true, zero-semantic-diff commit gives the walk almost nothing but the commit message to seed from. Adjustment: when docsOnly=true AND the diff stat shows only *.json paths, widen the walk's seed to include the referencing script's identifier tokens (e.g. scripts/milestones/*.mjs) rather than message text alone, so pure-data commits inherit their producer script's semantic home. Experiments: Experiment: re-run this commit's walk seeded with tokens from whichever script last wrote data/milestones-progress.json (grep scripts/ for writeFileSync.*milestones) instead of commit-message tokens only; measure whether offPct drops below 15 on replay · Experiment: add a docsOnly+all-json-diff detector to the receipt pipeline and tag those commits with a 'data-asset' flag in the receipt row; measure what fraction of the 20 out-of-lane 🎭 performer events in 14 days carry this flag · Experiment: lower prewalkOverlap floor alerting — flag any commit with prewalkOverlap<10 (vs litIntent/litReality>120) as 'seed-starved' before the walk runs, and report it as a distinct receipt category rather than a normal lane miss

     Measured outcome (deterministic, not self-graded): room mean offPct 17→50 over 4 commits since

     An unowned data-asset lane inflates this room's apparent loss ratio with a false claim — the fix is a one-line glob addition (near-zero cost) that immediately removes a structurally unfixable false-positive from the 20-event/14-day count, restoring the room's true premium signal.

     Open the signed attestation →

101. 2026-09-03 · 🎭 performer · offPct 43 · structural

     post(blog): the envelope needs a countable edge — co-opt cook of the agent-gym video, v1

     Why it left the lane: The commit's actual files (docs/05-content/blog/cook-rounds/*.claims.md, *.predictions.md, src/content/blog/2026-09-03-*.mdx) are the output of the mdx-cook skill's new sidecar/blog pipeline. None of those paths match ANY glob in 🎭 performer's owned surface (docs/decks/, insure-rounds/, deck/, event-*/, fundraising/, munich-re-*, over-exposure-*, demo-*). The commit message correctly lists Relevant-Rooms as 🎭 performer, 🎤 voice, ☕ network — but the receipt's reality walk saturates to litReality=144/144, the entire lattice, against litIntent=130/144, so 43% of a fully-saturated reality mass reads as outside whatever narrow lane the attribution script pinned to performer. The diff's own vocabulary ('premium/loss-ratio', 'underwriter', 'parametric trigger', 'coverage resolution', 'insurability') is genuinely performer-adjacent (insure-rounds pitch language) even though it is filed as a blog post, which is exactly the kind of legitimate cross-room content CLAUDE.md's Story/Relevant-Rooms trailer exists to describe.

     The semantic pull: Insurance-pitch terms embedded in the blog prose ('policy clause', 'parametric trigger', 'coverage resolution', 'premium', 'underwriter') pull the reef placement toward performer's real semantic territory (insure-rounds decks), while the physical path (src/content/blog/, docs/05-content/blog/cook-rounds/) has zero representation in ANY of the three declared Relevant-Rooms' owned-surface globs — so the walk has nowhere legitimate to land and spreads to full saturation instead.

     Intent–reality gap: The commit declares itself blog content correctly co-authored across performer/voice/network (Relevant-Rooms trailer, Story explains the cross-terminal cook). The receipt shows the reality walk did not settle into a bounded lane at all — it lit all 144 cells (full saturation) against an intent lane of 130 — so the declared cross-room authorship has no owned-surface glob to resolve against, and the detector reports maximal spread as 43% off-lane.

     • Add `src/content/blog/**` and `docs/05-content/blog/**` to the owned-surface glob list for 🎤 voice (or whichever room .workflow/rooms/*.html designates as the content/blog owner) — currently absent from every room's declared surface
     • Extend intervention-fire.sh's room resolution to union-match across all Relevant-Rooms-listed rooms, per the f9e88143b fix, and verify it actually reaches blog paths (this event suggests the union-match either isn't wired yet or the target room's glob set still has the gap)
     • Add a periodic/CI check (extending the f08ec5875 finding) that walks every path touched by the last N commits and flags any path matching ZERO room's owned-surface glob across all of .workflow/rooms/*.html
     • Re-run this commit's ballistic walk with a capped --max-depth to test whether litReality=144 (full saturation) is a depth artifact of a long multi-topic post rather than a genuine drift signal
     • Document in CLAUDE.md's mdx-cook / blog pipeline section which room owns docs/05-content/blog/cook-rounds/ sidecars, since that directory is new output from the mdx-cook skill and was never assigned an owner

     What we learned: f9e88143b already found that Relevant-Rooms is ignored in favor of Originating-Terminal; this event shows the escalation — even a properly union-resolved Relevant-Rooms match would STILL fail here, because the mdx-cook skill's new output directory (docs/05-content/blog/cook-rounds/) and the blog content path itself have never been assigned to ANY room's owned-surface glob. A brand-new content pipeline's output directory is invisible to the room-ownership model the moment it ships, not just when orchestration state files are shared (f08ec5875) — every new skill that writes to a new path needs its owning-room glob added in the SAME commit that ships the skill, or its first real commit is guaranteed to misfire regardless of attribution logic quality.

     The missing guard (the deliverable): A guard (e.g. tests/ops/no-ownerless-committed-paths.test.mjs) that, for every file changed in a commit, asserts at least one room in .workflow/rooms/*.html declares an owned-surface glob matching it — run in CI/post-commit, failing loudly (or logging UNOWNED, never silently attributing) the first time a new pipeline's output directory ships with no room claim.

     Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. Saturated, not right-sized: litReality=144/144 is the maximum possible fill of the 144-cell lattice — the walk found signal everywhere, which carries no discriminating power between in-lane and out-of-lane. Concrete adjustment: re-run with a capped --max-depth (e.g. ply 2) on this same commit and check whether litReality drops meaningfully below 144; if it does, saturation is a depth artifact for long multi-topic docs-only commits and the default depth should scale down for docsOnly=true commits. Experiments: Re-run pmu-onchip --ballistic --max-depth 2 against sha fdde61d21 and compare litReality (144 at default depth) to confirm whether capping depth un-saturates the walk for long-form multi-topic blog commits · Build the ownerless-path scan (per the missing_guard) against the last 90 days of commits and count how many off-lane events are pure glob-coverage gaps versus genuine drift, to size how much of the room's 19 recent out-of-lane events are this same class · Recompute this commit's offPct using union-of-Relevant-Rooms-owned-globs attribution instead of Originating-Terminal-only, and diff against the current 43% to directly measure the fix's effect on this exact event

     Measured outcome (deterministic, not self-graded): room mean offPct 17→20 over 8 commits since · dictionary ablation saved 1217B

     An ownerless-path misattribution inflates this lane's measured loss-ratio with drift the room never actually caused; premium restores the moment the blog-content glob is assigned to its real owning room and the attribution script matches against the full Relevant-Rooms union instead of the originating terminal alone.

     Open the signed attestation →

102. 2026-09-02 · 🎤 voice · offPct 18 · mistagged

     refactor(deck): rename /deck/seed to /deck/raise (v1.2-r1)

     Why it left the lane: Every semantic token in the diff — 'SeedDeckClient'/'RaiseDeckClient', 'DECK_VERSION', 'EXPOSURE', 'LICENCE_COST', 'FLEET', '/deck/seed'→'/deck/raise', the 'raise' fundraising vocabulary — is fund/pitch-deck content. 🎤 voice's own coordinate-lock (terminal-voice.html) declares archaeology.paths=["src/content/blog/*","books/*","public/book/*"] and explicitly lists 'Demo logistics' under 'ignores'. None of that overlaps src/app/deck/**, so the reality walk (117 lit cells, vs 69 on the intent side) fanned out into fund/deck vocabulary the voice lane never declared, and 18% of it landed outside any lane the receipt could anchor to voice.

     The semantic pull: 'raise'/'seed'/'deck'/'DECK_VERSION' and the dollar-exposure identifiers (EXPOSURE, LICENCE_COST, AUTHORITY_EACH, FLEET) are the pull — they are the literal owned-surface keyword ('deck') that 🎭 Performer's room page displays under A3 Strategy.Fund, not vocabulary voice's blog/book lane contains at all.

     Intent–reality gap: The commit message claims a narrow, self-contained rename ('route, client component, metadata title and guard all follow... guard re-run green') and that claim is true — but it is filed under 🎤 voice (Originating-Terminal, deterministic from $TERM_PROGRAM=Terminal.app) with Relevant-Rooms voice/architect/navigator, when the content is 100% fund/pitch-deck work that 🎭 Performer's own room page names as owned ('· deck' under A3 Strategy.Fund, prose: 'owns the decks').

     • Add "src/app/deck/**" and "docs/decks/**" to alacritty-performer.html's archaeology.paths (currently ["src/app/tesseract/*","docs/05-content/tesseract/battle-cards/*","public/"]) so the machine-readable glob matches the room's own displayed 'Owned: deck' label
     • Audit the other ~11 deck-family commits in this series (dentry/seed/grip/raise forks, per Performer's own recent-commits list which already shows 3 feat(deck) entries) for the same voice-vs-performer mistag pattern
     • Confirm whether the post-commit Gemini Relevant-Rooms suggester reads archaeology.paths at all, or free-guesses from the commit message/diff text with no glob check — this determines whether the config fix alone closes the gap or the suggester prompt also needs the glob surfaced to it
     • Cross-check src/app/deck/dentry, src/app/deck/grip for the same room-attribution gap now that the pattern is confirmed

     What we learned: A room's DISPLAYED 'Owned' list (the prose/HTML a human reads) and its machine-readable archaeology.paths glob (what any scorer actually matches against) can silently diverge — Performer's page has said 'deck' under Owned since at least the 2026-08-24 docs/decks/ claim, but the glob array was never updated to include it, so nothing reading the glob (rather than the prose) could ever route deck commits there. This is the NEVER-WRITE-STATE-INTO-A-GENERATED-FILE class applied to room ownership: the display and the data source drifted apart and nothing caught it because nothing diffs them.

     The missing guard (the deliverable): A guard that, for every commit, resolves each Relevant-Rooms entry's room file (.workflow/rooms/*.html), parses its archaeology.paths glob, and asserts at least one changed file matches — failing loudly (not just recording offPct) when a listed room has zero glob overlap with the diff. This commit is exactly the fixture: pre-fix, voice/architect/navigator all show zero overlap against src/app/deck/**.

     Detector probe: signal attribution — gzip reef 20% · ballistic walk 80%. Under-sized for this file shape: RaiseDeckClient.tsx is a rename touching 3 lines of code inside a ~700-line file with dense header commentary, so whole-file reingest inflates litReality far past the actual semantic delta (6 real line changes) and the walk truncates before resolving. Adjustment: for rename-class diffs (diff stat shows renames, low insertion count relative to file size), restrict reality ingest to the changed hunks plus immediate context rather than the full file. Experiments: Re-run the ballistic walk on 137192646b restricted to diff-hunk lines only vs. current whole-file reingest; measure whether offPct drops toward 0 — confirms whether comment-heavy whole-file ingest, not room mistagging, is doing extra damage here · Add src/app/deck/** + docs/decks/** to Performer's archaeology.paths, re-render this commit's receipt attributed to 🎭 performer instead of 🎤 voice, and check whether offPct falls under the 15 threshold — isolates how much of the 18% is pure room-mistag vs. genuine walk noise · Raise the walk's --max-depth budget specifically on walkTruncated=true events and measure whether litReality/litIntent converges below ~1.3x without exceeding the per-commit wall-clock budget

     Measured outcome (deterministic, not self-graded): room mean offPct 9→27 over 6 commits since · dictionary ablation saved 699B

     A correctly-scoped, guard-green rename filed under the wrong room's lane still books as loss experience against 🎤 voice (now 99 out-of-lane events in 14 days), inflating that lane's apparent drift rate with attribution noise rather than real content risk; premium on voice's lane is restored by fixing Performer's owned-surface glob so deck commits stop being scored against a lane that never claimed them.

     Open the signed attestation →

103. 2026-09-02 · 🧭 navigator · offPct 32 · structural

     feat(gcal-agent): charge a CBP buffer on international-origin connections, not just the hotel case

     Why it left the lane: The diff's own mass is dominated by strings that name OTHER domains than 'travel scheduling correctness': 'CBP', 'customs', 'preclearance', 'TSA re-screen', 'bag reclaim-and-recheck', 'jurisdiction' pull toward Operations.Grid/Loop/Flow (regulatory/logistics-flow topology) and Tactics.Deal/Speed (exchange, throughput, velocity) rather than staying inside a scheduling-check lane. The CITY table redesign (CITY_ZONE → CITY with {zone, cc, pre}) and the three-tier origin-resolution protocol (STATED/INFERRED/UNDECIDED) read, lexically, like a compliance/jurisdiction-classification engine, not a calendar-diff utility — that vocabulary is exactly what a gzip-reef seed keys on, and it is real content in the diff, not noise.

     The semantic pull: Literal strings dragging placement: 'CBP', 'customs', 'preclearance', 'PRECLEARANCE IS NOT A FOOTNOTE', 'jurisdiction', 'bag reclaim-and-recheck', 'TSA re-screen', the new `cc` field and `pre: true` markers in the CITY table, and 'STATED/INFERRED/NOTHING' as a three-way resolution protocol — this is the shape of a regulatory-flow classifier (pulls toward C1/C2/C3 Operations.Grid/Loop/Flow — topology, velocity, throughput) and a negotiation/exchange model (pulls toward B2.Tactics.Deal — 'exchange' labels on 3 of the top reds) more than it reads as B3.Tactics.Signal git-discipline work.

     Intent–reality gap: The commit claims a narrow, mechanical extension ('charge a CBP buffer on international-origin connections') scoped to one script (travel-conflicts.mjs) and one guard test — a same-lane, same-file feature addition. The receipt measured the actual TEXT mass as spreading across 31 out-of-lane cells covering topology/velocity/throughput/exchange semantics, because the diff's prose (comments + city/jurisdiction table) is dense enough to read as a jurisdiction/logistics-classification system rather than a scheduling arithmetic tweak.

     • Extract CITY table + cityInfo()/customsOnArrival()/selMatches() into scripts/gcal-agent/lib/customs-jurisdiction.mjs, imported by travel-conflicts.mjs
     • Add a one-line module header on the new file stating its sufficiency contract: 'answers only: does this city require a CBP buffer on arrival — not a general geo/jurisdiction service'
     • Update the commit-message template for gcal-agent feature commits to name the lookup table's bound explicitly (see spec_rewrite)
     • Add a CLAUDE.md-adjacent note under the room's owned-surface doc (if one exists for gcal-agent) naming CITY/jurisdiction data as belonging to this narrow scheduling-check lane, not a shared service
     • Cross-check tests/gcal-agent/travel-conflict-check.test.js still asserts against the extracted module path if the file_movement lands

     What we learned: A commit can be entirely correct, single-file, and test-guarded (both-directions red/green, per THE HARD RULE) and still trip the lane threshold — because the RECEIPT keys on the diff's comment/prose density, and a well-documented feature (this diff explains preclearance, jurisdiction, and a three-tier resolution protocol in ~120 lines of comments) generates enough 'regulatory/logistics' vocabulary to out-mass the 'scheduling check' vocabulary in its own file. Thorough in-file documentation is itself a source of lane drift risk when the documented subject (customs law) is lexically distant from the file's stated domain (calendar arithmetic).

     The missing guard (the deliverable): A pre-commit advisory (WARN-ONLY, per the STANDING TRIPWIRES pattern) that runs the same gzip-reef placement locally before commit and prints the top-3 predicted off-lane cells with their pulling terms — so a room can see 'this commit's prose is about to read as Operations.Grid/Tactics.Deal, not B3.Tactics.Signal' BEFORE it lands, and can choose to extract/rename rather than discover it after the receipt ships.

     Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Right-sized but asymmetric: litReality (137) exceeding litIntent (102) by 34% at only 2 plies suggests the reality corpus (comment-dense diff) is lighting more cells per ply than the intent corpus (terse commit-message claim) can match — not saturation (fill 0.98% is nowhere near the 70% ceiling) but an INPUT-MASS imbalance. Adjustment: increase the intent corpus's effective mass by feeding the FULL commit message body (already ~500 words) rather than a summarized claim, so litIntent has a fairer chance to cover the same ground litReality does. Experiments: Re-run this exact commit's placement with the commit message's full prose body (not a trailer-stripped summary) as the intent corpus, and measure whether offPct drops — tests whether intent-corpus truncation, not real drift, explains part of the 32% · Ablation: strip the PRECLEARANCE/customs-jurisdiction comment block from the reality corpus only (keep the code) and re-run the walk; if offPct falls sharply, the comment PROSE (not the code structure) is the primary drift driver, which would argue for a comment-density cap or a doc-vs-code split in the reality ingest · A/B the connectivity lattice's ply depth (2 vs 3) on this same commit and log litReality growth rate — if litReality keeps expanding past ply 2 disproportionately to litIntent, that's evidence the lattice under-connects intent-side terms and the fix is on the intent-ingest side, not the walk

     Measured outcome (deterministic, not self-graded): room mean offPct 28→12 over 4 commits since · dictionary ablation saved 933B

     A single well-guarded, both-directions-tested feature commit tripping offPct=32 nudges this lane's loss-ratio up on a false-positive-shaped event (the code is correct, per the firing-test discipline) — premium is restored by shipping the pre-commit lane-preview advisory (missing_guard) so future dense-comment commits in this room self-correct before the receipt fires, rather than by disputing this one reading.

     Open the signed attestation →

104. 2026-09-02 · 📐 architect · offPct 19 · structural

     fix(voice): prep's Ava briefing has never once attached — two stacked defects, both guarded

     Why it left the lane: The reality walk lands squarely on scripts/say-to-mp3.sh and the two new tests/voice/*.test.sh files — vocabulary like 'lame --quiet', 'ps -o etimes=', 'pgrep -x say', 'SAY_MP3_ARGS', 'Ava (Premium)', 'ffprobe' has zero gzip-NCD overlap with 📐 architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, CLAUDE.md). prewalkOverlap=6 is low relative to litIntent=123/litReality=128, meaning the reef seed itself found little shared vocabulary between architect's declared lane and this commit's actual bytes before the walk even ran — the walk then dutifully lights up 128 cells in a lattice neighborhood (🎤 voice) that was never in the intent seed at all.

     The semantic pull: 'lame -V4' → speech-encoding fix, 'ps -o etimes= -p $pid | tr -dc 0-9' → the reap-race fix, and both new test files under tests/voice/ — every one of these strings is native 🎤 voice-domain content; none intersects architect's strategy/doc glob, which is exactly why offPct reads 19 against a lane built from Originating-Terminal alone.

     Intent–reality gap: The commit's own trailers already declare the true intent correctly — 'Relevant-Rooms: 🎤 voice, 📐 architect, 🎩 operator', voice named first — but the receipt's offPct is computed against the Originating-Terminal lane only (architect), so a commit that self-labeled its cross-room nature accurately still reads as 19% out-of-lane drift.

     • Grep for the three unshipped room-drift-<room>-check.sh specs (architect/navigator/laboratory) referenced in ff09774e5 and consolidate into one file rather than a fourth parallel spec
     • Wire the consolidated classifier into whatever module currently reads Originating-Terminal for offPct computation
     • Confirm scripts/say-to-mp3.sh and tests/voice/** are present in 🎤 voice's owned-surface glob in .workflow/rooms/voice.html (add if missing)
     • Backfill: recompute offPct for the last 189 architect out-of-lane events using the union-lane rule and report what fraction were Relevant-Rooms-divergence false positives vs genuine drift
     • Add the binding sentence to CLAUDE.md's commit-trailer section (see spec_rewrite) in the same commit as the classifier
     • Ship tests/pmu-simulator/room-lane-uses-relevant-rooms.test.mjs before closing this out — see missing_guard

     What we learned: This is not a new mechanism — it is the SAME root cause ff09774e5 already diagnosed, now recurring a fourth time with zero ships: three rooms independently spec'd the identical fix and none merged it, and this event proves the gap is still live. The structural fact this reveals is one level up from the detector bug itself — the repo's self-heal loop can correctly diagnose the same defect repeatedly (SELF-IMPROVING FIRES' 'diagnose the root' step is working) without ever completing the 'ship the prevention' step, which means diagnosis-without-convergence is now the dominant failure mode of the intervention system itself, not of any individual commit.

     The missing guard (the deliverable): tests/pmu-simulator/room-lane-uses-relevant-rooms.test.mjs — feed the classifier a synthetic commit with Originating-Terminal: architect, Relevant-Rooms: voice, and changed files matching voice's owned-surface glob; assert it does NOT trip the offPct threshold. Pair it with a genuinely mistagged commit (no Relevant-Rooms glob match at all) and assert THAT one still trips — both directions, per EVERY GUARD SHIPS WITH A FIRING TEST.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/litReality sit at 123/144 and 128/144 — roughly 85-89% of the lattice lit on both sides. That's close to saturated, not shallow: near-total lit coverage compresses the signal-to-noise on 'off-lane', because almost everything is lit somewhere. Concrete adjustment: raise the lit-threshold (or lower --max-depth) so the walk stops propagating once a ply's marginal new-cell yield drops below some floor, which should sharpen the distinction between genuine off-lane mass and saturation noise. Experiments: Seed the reef dictionary with voice-domain terms tagged to 🎤voice's coordinate (lame, ffprobe, pgrep, etimes, Ava Premium) and re-run this commit's receipt — measure whether offPct drops purely from vocabulary coverage, isolating reef-seed contribution from lane-logic contribution · Recompute offPct for the last 189 architect out-of-lane events under the union-lane rule (Originating-Terminal ∪ Relevant-Rooms globs) and report what fraction resolve under the 15% threshold — this directly tests the structural-failure hypothesis at scale · Bucket the 189 events by prewalkOverlap value and check whether prewalkOverlap<10 predicts 'mistagged/structural' class specifically — if it does, the detector could auto-flag likely-false-positive drift without an LLM pass at all

     Measured outcome (deterministic, not self-graded): room mean offPct 11→10 over 5 commits since · dictionary ablation saved 668B

     A false-positive lane flag like this degrades trust in the offPct signal itself — it inflates 📐 architect's apparent loss-ratio with zero real cross-lane risk — and is restored only by shipping the Relevant-Rooms-aware classifier so offPct fires exclusively on genuine boundary failures rather than correctly-declared cross-room work.

     Open the signed attestation →

105. 2026-09-02 · 🎤 voice · offPct 23

     feat(deck): raise v1.2-r7 — the pick-a-lane pass: frames 11-17 scannable, decidable on the front

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 21→27 over 5 commits since

     Open the signed attestation →

106. 2026-09-02 · 🔒 vault · offPct 23 · mistagged

     feat(lens): the receipt bands the walk by colour — greens, ambers, reds

     Why it left the lane: The diff contains zero legal/patent/vault vocabulary — no 'patent', 'claim' (legal sense), 'NDA', 'license', 'data-room'. It is pure PMU-lens engineering: 'fenceDist', 'toleranceBand', 'shortlex-coords.mjs', 'lens-encircled-png.mjs', 'walkProvenance', 'encircledBands'. There is no ambiguous vocabulary dragging the walk toward vault's coordinates — the mistag isn't semantic at all, it's mechanical: the commit trailer names Originating-Terminal: vault (the physical terminal the operator typed in) and Relevant-Rooms echoes vault first even though vault's owned globs (docs/01-business/patents/, docs/legal/, data-room*/, docs/incidents/) match zero of the four changed files.

     The semantic pull: None of the changed paths (scripts/pmu/lens-encircled-png.mjs, scripts/pmu/prompt-lens.mjs, scripts/pmu/shortlex-coords.mjs, tests/pmu-simulator/encircled-bands-by-colour.test.mjs) intersect vault's declared surface at all — the 'pull' toward vault is zero real overlap, which is itself the anomaly worth naming: a mistag with no semantic cause is a pure attribution-mechanism bug, not a content ambiguity.

     Intent–reality gap: The commit message and trailers claim relevance to 🔒 vault (plus navigator, architect); the receipt measured 127–134 lit walk cells entirely inside PMU/ShortLex lens machinery with prewalkOverlap=3 against whatever corpus vault's lane seeds — the receipt is scoring engineering work against a patents/legal intent lane the commit never actually claims to serve.

     • Locate the script/hook computing the Relevant-Rooms trailer (grep -rl "Relevant-Rooms" scripts/ hooks/) and patch it to exclude any room with zero owned-surface glob matches instead of falling back to Originating-Terminal.
     • Add scripts/pmu/** and tests/pmu-simulator/** to the owned-surface glob of the room that should cover PMU/lens engineering (architect, given it's already listed second) in .workflow/rooms/*.html.
     • Add tests/rooms/relevant-rooms-glob-match.test.mjs asserting every room in a commit's Relevant-Rooms trailer has ≥1 real glob match against the changed files.
     • Re-audit vault's 20 out-of-lane events from the last 14 days to separate genuine cross-lane vault work from this exact mistag mechanism before trusting the room-mean trend.
     • Name scripts/pmu/**'s owning room explicitly in CLAUDE.md's PMU section, closing the 'coordinate owns no code surface' gap per DELEGATE-TO-SUBTASKS.

     What we learned: scripts/pmu/** — one of the most actively-churned subsystems in the repo — is not covered by ANY of the 9 canonical rooms' owned-surface globs. Prior lessons fixed which field to read and caught the Originating-Terminal fallback firing; this event shows WHY the fallback keeps firing for PMU work specifically: there is no correct room for the glob-matcher to find, so every PMU commit is structurally guaranteed to mistag to whichever terminal happened to be open, regardless of algorithm correctness.

     The missing guard (the deliverable): tests/rooms/relevant-rooms-glob-match.test.mjs — for a real historical commit, assert every room named in Relevant-Rooms has ≥1 changed-file match against its declared glob in .workflow/rooms/*.html; fails loudly instead of silently re-legitimizing Originating-Terminal with zero matches.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent(134) and litReality(127) are closely matched in fill (~5% spread) so the walk isn't obviously saturated or too shallow in volume, but walkTruncated=true means it hit its time/depth budget before fully resolving. Concrete adjustment: re-run this sha with --max-depth+1 (or one notch more timeout budget) to confirm offPct is stable and not partly an artifact of early truncation. Experiments: Re-seed the intent corpus for vault commits using ONLY docs/01-business/patents/**, docs/legal/**, data-room*/**, docs/incidents/** (never whole-repo docs), then re-render this sha's receipt and confirm prewalkOverlap stays near 0 — proving true lane-mismatch rather than a corpus-ingestion defect. · Add scripts/pmu/** as an owned glob for architect, recompute this commit's Relevant-Rooms, and re-render the receipt against the corrected intent lane — measure whether offPct falls toward the architect room's mean once scored against the right lane. · Strip Originating-Terminal auto-inclusion from Relevant-Rooms for the last 14 days of vault commits and recompute the room-mean offPct trend, isolating how much of the '20 out-of-lane events' figure is this mechanism versus genuine cross-lane work.

     Measured outcome (deterministic, not self-graded): room mean offPct 16→1 over 4 commits since · dictionary ablation saved 831B

     This event is the 20th vault out-of-lane tick in 14 days from a mistag mechanism, not genuine cross-lane risk, so it's inflating vault's apparent loss-ratio on paper; premium correctness is restored once scripts/pmu/** gets a real owning-room glob and the Relevant-Rooms trailer stops crediting Originating-Terminal with zero-match rooms.

     Open the signed attestation →

107. 2026-09-02 · 🎭 performer · offPct 31

     deck(raise): name the category — semantic parametric triggers

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 12→17 over 6 commits since

     Open the signed attestation →

108. 2026-09-02 · 🎤 voice · offPct 19 · mistagged

     feat(deck): raise v1.2-r3 — the fee stack: ThetaDriven paid at mint, at transfer, at settlement

     Why it left the lane: The entire diff lives in two files — src/app/deck/raise/RaiseDeckClient.tsx and tests/deck/raise-deck.test.mjs — and every added string is investor/pricing vocabulary: 'Mint · $20/agent-yr', 'Resale · 10% of every re-registration', 'Roll-up · every book is licensed agents', 'Settlement · paper pays only against the named key', 'financialization is what is licensed', 'keep the spread'. None of that touches books/tesseract/chapters, src/content/blog, decentralizedtalk, or 6needs — voice's actual owned surface. The reef has no book/blog-shaped mass to grip in this diff, so the walk's litReality (116) lands on whatever GTM/pricing-adjacent coordinates the gzip-NCD lexicon nearest-matches ('licence', 'settlement', 'fee', 'financialization' cluster with strategy/vault vocabulary), producing offPct=19 against a claimed voice lane the diff never actually enters.

     The semantic pull: Deck-building lexicon (mint/resale/rollup/settlement, licence, financialization) has zero overlap with the prose/narrative vocabulary that defines voice's lane (book chapters, blog craft, 6needs). No book/blog file appears anywhere in the diff — the reef mass is 100% architect/vault-flavored pricing content wearing a voice-terminal trailer.

     Intent–reality gap: The commit message accurately describes what it did (assembled a fee-stack diagram on the raise deck) and even lists Relevant-Rooms as voice, vault, architect — but the receipt's lane is keyed to Originating-Terminal (🎤 voice), a room whose owned surface never included src/app/deck/** or docs/decks/**. The work is correct; the lane it's measured against is the wrong one.

     • Add `src/app/deck/**` and `docs/decks/**` to 📐 architect's owned-surface glob in `.workflow/rooms/architect.html` (Relevant-Rooms already names architect for this commit — the owned-surface declaration should match)
     • Note in CLAUDE.md's deck skill / SOVEREIGN OUTREACH section that deck-building work is owned by 📐 architect regardless of which terminal (Originating-Terminal) physically ran the commit
     • Audit the room's 101 out-of-lane events over 14 days for a `src/app/deck/**` / `docs/decks/**` file-path signature — if a large share share this signature, the fix here retroactively explains most of the backlog rather than being a one-off
     • Add a guard asserting the receipt's room-lane computation prefers owned-surface glob match over raw Originating-Terminal when the two diverge on a commit whose files are 100% outside the Originating-Terminal room's surface

     What we learned: The room taxonomy has no owning room for `src/app/deck/**` — deck-building work is currently always attributed to whichever terminal ($TERM_PROGRAM) happened to be open, not to the room whose owned-surface glob actually matches the changed files. This is a structural gap in the 9-room ownership map, not a per-commit slip, and it will keep firing on every deck commit made from a non-architect terminal until the glob is added.

     The missing guard (the deliverable): A test asserting: for any commit whose changed files are ⊆ a room's owned-surface glob (e.g. architect's), the Originating-Terminal trailer's room must either match that owning room or the receipt's lane computation must fall back to the glob match rather than the terminal — i.e. `tests/rooms/lane-prefers-owned-surface-over-terminal.test.mjs`.

     Detector probe: signal attribution — gzip reef 30% · ballistic walk 70%. walkTruncated=true with litReality(116) < litIntent(129) suggests the walk hit its time/depth budget before reality fully expanded to match intent's coverage — under-filled, not saturated. Concrete adjustment: re-run this sha's walk with --max-depth+1 (or a larger step budget) and check whether litReality converges toward 129; if it does, the offPct is partly a truncation artifact rather than a true lane mismatch. Experiments: Re-score this commit's lane using owned-surface-glob match instead of Originating-Terminal and diff against the current offPct=19 to measure how much of the drift is attribution-artifact vs real semantic distance · Re-run the ballistic walk on 8ec956a69 with an incremented --max-depth and compare litReality against the current 116 to test the truncation hypothesis directly · Retroactively re-score the room's other 100 out-of-lane events from the last 14 days against a `src/app/deck/**`/`docs/decks/**` path filter to measure what fraction of the 101-event backlog shares this exact misattribution signature

     Measured outcome (deterministic, not self-graded): room mean offPct 16→48 over 4 commits since · dictionary ablation saved 976B

     A mistagged lane inflates voice's apparent loss ratio with content it was never underwritten to cover; premium on the voice lane is restored the moment deck-path commits score against architect's owned surface instead of the physical terminal that happened to run them.

     Open the signed attestation →

109. 2026-09-02 · 📐 architect · offPct 28 · structural

     docs(ticker): correct the stale @domain header, name the email-keyed aggregation

     Why it left the lane: The diff is a comment-only edit to src/components/RenderTicker.tsx, a file that matches none of architect's owned-surface globs (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The comment text itself is dense architecture-planning prose naming real lib symbols — 'aggregateDomains', 'computeTicker', 'src/lib/render-ticker-core.mjs', 'email.split(\'@\')[1]' — plus a proposed data shape '{domain, weight, at}' and 'a per-source adapter' keyed on DOMAIN. It also uses 'a receipt or a licence has an org, not an inbox' twice, vocabulary that is heavily overloaded elsewhere in this repo's pricing/insurance/attestation apparatus. None of that vocabulary belongs to a docs/strategy lane; it reads exactly like real lib/architecture and pricing-domain work, so the walk placed 28% of lit reality mass in those adjacent coordinates.

     The semantic pull: litIntent=116 vs litReality=139 with prewalkOverlap=4 (only ~3.4% shared before the walk even ran) shows the divergence starts at the reef seed, not just the walk: the intent corpus (architect's docs-strategy lane + this commit's own claim 'comment-only, no behaviour moved') seeds narrow, while the reality corpus (the diff text) seeds into lib/architecture and receipt/licence-adjacent coordinates the moment it's placed. walkTruncated=true while litReality is still ahead of litIntent means the reality side was still expanding when the budget ran out — the pull is real and unresolved, not an artifact of a finished walk.

     Intent–reality gap: The commit claims to be 'Comment-only change to the render path; no behaviour moved' — a minimal, docs-flavored edit. The receipt measured 28% of the lit reality mass landing outside architect's lane, because the comment itself is a full architecture proposal for a different file (render-ticker-core.mjs) that the detector cannot distinguish from an actual code change to that file.

     • Extract the 'aggregation is not yet source-agnostic' paragraph from RenderTicker.tsx into docs/03-product/ticker-source-agnostic-aggregation.md; leave a one-line pointer comment in the component
     • Ship the room-drift-lane-check.sh that reads the Relevant-Rooms trailer (not just Originating-Terminal) into the intent-lane classifier — spec'd independently by architect, navigator and laboratory per ff09774e5, still unshipped
     • Add a pre-commit check flagging comment blocks over ~10 lines containing architecture-planning vocabulary ('adapter', 'shape that absorbs', proposed function signatures/data shapes) in a file outside the committing room's owned-surface globs
     • Disambiguate 'receipt'/'licence' in the gzip reef dictionary by file-type/room context so a UI-comment's incidental use of those words stops dragging placement toward the pricing/insurance lattice
     • Re-render this commit's receipt after the doc-extraction to confirm offPct drops under the 15 threshold
     • Note in .workflow/rooms/architect.html's owned-surface section that planning prose authored from architect's terminal but living inside another room's file must be filed to docs/03-product/, not left inline

     What we learned: A commit can be textually true ('comment-only, no behaviour moved') and still trip lane drift, because the detector measures the SEMANTIC WEIGHT of the prose, not whether bytes execute — a comment that describes a future refactor in enough structural detail (function names, data shapes, keying rules) reads to the walk exactly like the refactor itself. 'No behaviour moved' is not the same claim as 'no architecture claimed', and only the second one bounds the lane.

     The missing guard (the deliverable): A structural-density check on comment blocks: any comment diff naming ≥2 concrete code symbols (function names, file paths, proposed data shapes) and exceeding ~8 lines, in a file outside the committing room's owned-surface globs, fails pre-commit unless it also contains a docs/ pointer — forcing architecture-planning prose to live where it's claimed to live.

     Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. litReality is still climbing past litIntent when walkTruncated=true fires — the reality side is under-budgeted for comment-dense commits, not saturated. Concrete adjustment: raise --max-depth (or the walk time budget) specifically when the reality corpus is prose-dense (comment/doc content) rather than applying a flat budget across code and comment-only commits. Experiments: Re-render 921b6c220 with 'receipt'/'licence' disambiguated by file-type context in the reef dictionary and measure whether offPct drops below 15 · Implement Relevant-Rooms-aware intent-lane seeding and re-render this same commit to test whether offPct changes when the lane includes performer+laboratory alongside architect · Ablate --max-depth from 116-budget to a higher ceiling on this commit's reality walk and check whether litReality stabilizes (true saturation) or keeps climbing (confirms truncation artifact)

     Measured outcome (deterministic, not self-graded): room mean offPct 17→5 over 8 commits since · dictionary ablation saved 859B

     Comment-only, reversible risk shouldn't move this lane's premium on its own, but a room already running 191 out-of-lane events in 14 days with a known, three-times-spec'd-and-unshipped classifier defect is unpriced tail risk — the loss ratio won't improve until the Relevant-Rooms-aware lane check actually ships.

     Open the signed attestation →

110. 2026-09-02 · 🎤 voice · offPct 44 · structural

     fix(deck): raise v1.2-r2 — three-reader monologue fixes: kill shots, arithmetic, definitions

     Why it left the lane: The commit message frames this as pure prose repair ("kill shots, arithmetic, definitions"), which is exactly the voice room's mandate. But the diff itself is dominated by JSX/TSX structural markup that never appears in voice's owned globs: `<strong className="text-[#241f18]">`, `<Link href="/pricing" className="underline decoration-[#a8320d]...">`, `<div className="flex h-full flex-col justify-center px-6 md:px-14">`, plus a second file of `assert.match(SEED, /DECK_VERSION.../)` and `assert.ok(/3,154 metered units/.test(flat)...)` test code. The reef's gzip-NCD placement reads those bytes as component/test-code mass, not blog/book prose mass, and the room's declared surface (`books/tesseract/chapters/*.md`, `src/content/blog/*.mdx`, `docs/05-content/blog/scratchpad/`) has no entry for `src/app/deck/**` at all — so every touched byte in this commit is structurally unrecognized by the lane the commit claims.

     The semantic pull: Two forces pulled the mass off-lane simultaneously: (1) the file path itself — `src/app/deck/raise/RaiseDeckClient.tsx` and `tests/deck/raise-deck.test.mjs` sit entirely outside any glob voice owns, so there is no lane for the reef to seed into even before reading content; (2) the payload is JSX-wrapped prose (`<p className=... >That ledger made capitalism...</p>`), so the actual sentence-level edits (the ones that ARE voice's job — 'a one-book record is the thing capital walks away from', the Lloyd's-market correction, the Hooper holding fix) are inseparably fused with markup and guard-assertion code that reads as architecture/test surface.

     Intent–reality gap: The message claims a content-editing pass (three blind readers fixing kill shots, arithmetic, and definitions in deck prose) — a voice-room task by every description. The receipt measured 44% of the lit mass landing outside the intent lane because the only place that prose can live is inside a React component + its guard test, neither of which the room's owned-surface glob recognizes; the claim and the file topology were never going to agree.

     • Extract the deck's frame/section content out of RaiseDeckClient.tsx into a dedicated content module (raiseDeckFrames.ts) with no JSX markup mixed in
     • Add `src/app/deck/**/*Frames.ts` to voice's owned-surface glob in `.workflow/rooms/voice.html`
     • Split tests/deck/raise-deck.test.mjs into a content-assertion suite (imports the frames module directly, owned by voice) and a render/structure suite (imports the component, owned by architect/builder)
     • Re-run the receipt against the split commit shape to confirm offPct drops under the 15 threshold for a pure content edit
     • Note in CLAUDE.md's deck-skill section that deck copy is voice-room content and deck component/test wiring is architect/builder, so future delegation doesn't re-conflate them
     • Seed the reef dictionary with the repeated className boilerplate strings (compression_dictionary_adds) and re-check whether prewalkOverlap rises

     What we learned: This is the first of four out-of-lane events for this room where the mechanism is nameable rather than facts-only, and it reveals a structural gap distinct from the prior three: pitch-deck prose has no content/markup separation anywhere in the repo — copy lives fused inside a TSX component tree, so any voice-room prose task on a deck is structurally forced to touch a non-owned code surface no matter how disciplined the edit is. The prior lessons (offPct trending 27→39, 0→36, 26→20) never diagnosed a mechanism because none of those commits, apparently, had this content/markup fusion named as the cause.

     The missing guard (the deliverable): A structural test asserting that any file under `src/app/deck/**` containing rendered prose paragraphs (JSX `<p>`/`<h2>` text nodes) has that prose sourced from an imported content module matching the room's owned glob, rather than inlined directly in the component — i.e. `tests/deck/content-separation.test.mjs` failing if RaiseDeckClient.tsx (or any deck client) contains literal prose strings longer than N characters directly in JSX rather than via import.

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. litIntent/walksIntent = 123/123 and litReality/walksReality = 124/124 — both sides are lighting up essentially every coordinate they walk to, i.e. fully saturated with zero discrimination margin. That is too shallow/undifferentiated for a commit this mixed (prose + JSX + test assertions): concrete adjustment is to increase `--max-depth` (or tighten the lit-threshold) specifically for deck-room commits so the walk can separate the prose-mass sub-cluster from the markup/test-mass sub-cluster instead of lighting both uniformly. Experiments: Re-render this exact commit's receipt against a scratch tree where the frame content has been pre-extracted into a standalone file, and measure whether offPct drops toward/under 15 — isolates whether content/markup fusion is really the driver · Seed the gzip reef dictionary with the repeated className boilerplate strings below and re-measure prewalkOverlap on the next 5 deck-room commits to see if it rises off the floor of 3 · Add a diff-shape pre-classifier (ratio of quoted-string-literal bytes changed vs total bytes changed) that tags a commit 'content-dominant' above ~70%, and A/B the reef seed weighting for content-dominant commits against the current uniform seeding to see if offPct on content-only deck edits falls

     Measured outcome (deterministic, not self-graded): room mean offPct 16→48 over 4 commits since · dictionary ablation saved 1293B

     An out-of-lane event this large (44% against a 15% threshold, on a room already running 100 such events in 14 days) reads as unmanaged scope creep and raises this lane's loss-ratio; premium is restored by the content/markup split plus the glob update, which gives future pure-copy commits a lane the reef can actually place them in.

     Open the signed attestation →

111. 2026-09-02 · 🧭 navigator · offPct 23 · mistagged

     feat(gcal-agent): diff hotel checkout against the same-day flight, so an itinerary that contradicts itself says so

     Why it left the lane: The diff is a single 371-line new module (travel-conflicts.mjs) plus a 256-line test file plus a 5-line hook into day-dump.sh. That's a full standalone timezone-resolution + arithmetic engine (wallToUtc, offsetMs, partsIn, CITY_ZONE table, regex parsers for flight/checkout text) landing under scripts/gcal-agent/ — a directory the navigator room does not own (room surface is scripts/open-room-session.sh, .thetacog/cache/room-runners/, openclaw/*, .workflow/cognitive-dashboard.html). The reef pulls toward B3.Tactics.Signal (×11, the strongest single magnet) and C1/C2/C3 Operations.Grid/Loop/Flow — exactly the shape of 'build a new deterministic checking/detection module with an inline arithmetic construct and a guard,' which is the git-discipline/build-and-guard lane, not the navigator's session-routing/thematic-scheduler lane. The commit trailers even self-report Relevant-Rooms as navigator/operator/builder — three rooms, one owning file surface.

     The semantic pull: Terms doing the dragging: 'latest safe checkout = departure − transit − bag-drop − boarding' (a construct, not a keyword — pulls to Operations.Grid/topology and Operations.Flow/throughput), 'Guard ships in the same commit and was watched going red' (pulls straight to B3.Tactics.Signal, the guard-discipline axis), 'mutating CUTOFF to 0 turns five of the nine subtests red' (Operations.Loop/velocity — a state-mutation-and-observe pattern), and the sufficiency-contract paragraph ('WHAT IT IS SUFFICIENT FOR... declines the undecidable half') which is verbatim AXIOM-1 vocabulary that this repo's reef has learned to route to the builder/signal lane regardless of which room wrote it.

     Intent–reality gap: The commit message frames this entirely as a navigator-room travel-itinerary fix ('the body to the place'), but the actual work — a from-scratch deterministic detector with a guard-discipline narrative, a sufficiency contract, and a bootstrap-case test methodology — is textbook builder-room content that happens to live under scripts/gcal-agent/, a directory none of the three claimed rooms' owned-surface globs include.

     • Add `scripts/gcal-agent/**` (or at minimum `scripts/gcal-agent/travel-conflicts.mjs` + its test) as an explicit owned-surface glob to whichever room's `.workflow/rooms/*.html` file is meant to own trip-logistics detectors — today it is unclaimed by all nine.
     • In future gcal-agent commits, order the Relevant-Rooms trailer by actual code shape (guard-bearing new module → builder first) rather than by which terminal was physically open.
     • Cross-link scripts/gcal-agent/travel-conflicts.mjs's own header comment to the room that now owns it, once assigned, so a future reader doesn't have to re-derive ownership from the reef.
     • Audit the other 5 out-of-lane events in navigator's last-14-days history (per the room stat above) for the same unclaimed-directory pattern before treating this as a one-off.

     What we learned: scripts/gcal-agent/ is a real, growing directory (day-dump.sh + now travel-conflicts.mjs + its test) that sits entirely outside every room's declared owned-surface glob list — the room-ownership map has a blind spot for an entire subsystem, not just this one file, so every commit into it will keep mistagging until the glob is added somewhere.

     The missing guard (the deliverable): A test asserting every file changed in a commit is covered by at least one room's owned-surface glob in .workflow/rooms/*.html — flag (warn-only, per the recontext skill's own contract) any changed path matching zero globs, so an unclaimed-directory drift like scripts/gcal-agent/ surfaces at commit time instead of being discovered by a downstream offPct spike.

     Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Asymmetric and truncated — litReality (143) outpacing litIntent (102) by 40% while walkTruncated=true means the walk was cut off before it could reconcile that gap, which inflates offPct artificially. Concrete adjustment: rerun this sha's walk with --max-depth bumped one ply higher specifically for commits whose diff introduces a brand-new top-level directory (structural novelty is a cheap, computable signal for 'give this walk more budget'), and compare offPct before/after to see if truncation alone explains part of the 23%. Experiments: Recompute this sha's receipt with --max-depth+1 and diff the resulting offPct against 23 — if it drops meaningfully, truncation (not real drift) explains part of this event, and the walk-budget-on-novel-directories heuristic above is validated. · Run the missing_guard's glob-coverage check against the last 90 days of commits and count how many out-of-lane events land in directories matching zero room globs — if it's a large fraction, the unclaimed-directory blind spot generalizes far beyond gcal-agent and the fix should be a repo-wide glob audit, not a single addition. · A/B the reef seed with and without prewalkOverlap's current 6-anchor result stripped, to see how much of the 23% offPct is attributable to the walk alone (isolating the reef_signal_pct/walk_signal_pct split empirically instead of by estimate).

     Measured outcome (deterministic, not self-graded): room mean offPct 23→32 over 4 commits since · dictionary ablation saved 1143B

     An unclaimed-directory mistag is a room-attribution loss, not a code-quality one — the guard itself is real and was watched red-to-green, so premium impact is limited to this lane's ownership-clarity signal; it's restored the moment scripts/gcal-agent/** gets a glob owner and the coverage guard ships.

     Open the signed attestation →

112. 2026-09-02 · 🎭 performer · offPct 26

     book(tesseract): Three Rivers One Name (ch6) + The Lights Were Already On (conclusion)

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 14→10 over 20 commits since

     Open the signed attestation →

113. 2026-09-02 · 🎤 voice · offPct 55 · structural

     fix(deck): raise v1.2-r9 — round-3 repairs: the pen goes to licensed carriers, every figure cited or cut

     Why it left the lane: No cached region narrative exists for bbea8e6e06, so the mechanism has to be read off the diff and the ownership globs directly. The three touched files — src/app/deck/raise/RaiseDeckClient.tsx, tests/deck/raise-deck.test.mjs, public/decks/raise-deck.pdf — sit entirely outside 🎤 voice's declared surface (books/tesseract/chapters/*.md, src/content/blog/*.mdx, docs/05-content/blog/scratchpad/, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/, scripts/voice-*-check.sh). The reef seeds on prose vocabulary from that glob; the reality corpus here is a TSX component whose surrounding mass is JSX/Tailwind boilerplate — `className="border-t-2 border-[#a8320d] pt-4"`, `text-[clamp(0.72rem,1.2vw,0.9rem)] uppercase tracking-[0.16em]`, the `DECK_VERSION = 'v1.2-r9'` constant bump, and a test file full of regex assertions (`/Licensed carriers and desks write/`, `/\$10\.5tn/`). prewalkOverlap=3 against litIntent=124 (≈2.4%) shows the mismatch starts at the SEED stage, before the walk even propagates: almost none of voice's prose-lane reef vocabulary is present in this file's surrounding structure, even though the literal prose being edited ("Carriers write the risk", "$10.5tn", "Cybersecurity Ventures") is exactly voice-craft content.

     The semantic pull: The actual content edits — killing the preachy self-selection sentence, attributing risk-writing to licensed carriers, citing Cybersecurity Ventures for $10.5tn, softening the Lloyd's claim — are voice-craft strings and would seed cleanly into voice's coordinate. But they are embedded as JSX children inside a component whose surrounding syntax (Tailwind arbitrary-value classes, `PCT_OF_EXPOSURE`/`LICENCE_COST` constant arithmetic, and a co-edited assertion-heavy .test.mjs) is engineering/QA vocabulary that pulls reality mass toward operations/tactics-flavored coordinates instead of the content lane the commit claims.

     Intent–reality gap: The message claims a content-only repair — 'round-3 repairs,' 'every figure cited or cut,' voice-craft language throughout — and Persona-Intent frames it as reader-facing prose grading. The receipt measured 55% of the lit reality mass landing outside the claimed lane, because the only way to make that prose edit was to touch a .tsx component and its paired assertion test — paths the voice lane was never declared to own.

     • Extract deck prose into src/app/deck/raise/copy.ts (or equivalent) and re-import into RaiseDeckClient.tsx
     • Add the new copy-file glob to voice's owned-surface declaration in .workflow/rooms/voice.html
     • Re-run the reef ingest so the new copy.ts content actually seeds the voice corpus (glob change alone doesn't backfill the reef)
     • Audit other TSX-embedded decks (the dentry fork referenced in the diff's header comments, any docs/decks/*.html equivalents) for the same copy-in-component tangle and apply the same split where they recur
     • Add a WARN-only pre-commit check (recontext-style) that flags when Originating-Terminal's room owns none of the commit's touched paths, carrying a named exception for the copy.ts pattern once it exists
     • Update tests/deck/raise-deck.test.mjs comment header to note the copy/structure split once file_movements land, so future readers don't re-tangle them

     What we learned: This repo already has a `/recontext` skill built exactly to compute room-vs-surface mismatches like this one, but nothing runs it automatically at commit time — the mismatch between a commit's Originating-Terminal room and its actually-touched paths is currently only discoverable by manually diffing the owned-surface globs against `git show --stat`, which is precisely the gap the drift receipt is meant to catch structurally instead of by hand.

     The missing guard (the deliverable): A WARN-only pre-commit (or post-commit, non-blocking) check that computes, for the staged/committed paths, whether any of them fall under the Originating-Terminal room's declared owned-surface globs — and if none do, prints the mismatch loudly (same posture as the WIP-claims and rules-for advisories) rather than staying silent until a drift receipt fires days later.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true with litIntent/litReality both under 150 on a 144-lattice budget suggests the walk hit its depth/ply cap before fully exploring connectivity from the TSX/JSX-heavy file — JSX has denser nesting than markdown, which can eat plies faster. Concrete adjustment: re-run this sha's walk with --max-depth increased by 2 plies and compare offPct; if it drops materially, the truncation (not just the seed mismatch) is inflating the 55% figure and the default depth should scale with file syntactic density, not stay fixed. Experiments: Re-render bbea8e6e06's panel with --max-depth +2 and record whether offPct falls, isolating truncation's contribution from the seed-mismatch contribution · After extracting copy.ts (code_rewrite) and updating the glob, re-render an equivalent future copy-only commit and measure whether offPct lands under the 15 threshold · Seed the voice reef corpus with a small hand-picked vocabulary list drawn from this diff (underwriter, carriers, attestation, settlement, Lloyd's, licensed) and re-measure prewalkOverlap on a held-out deck-copy commit to see if it climbs off the near-zero floor

     Measured outcome (deterministic, not self-graded): room mean offPct —→15 over 4 commits since · dictionary ablation saved 927B

     A deck-copy commit that must touch a non-owned .tsx path to make a legitimate content fix keeps this lane's loss-ratio elevated every time it recurs (105 prior out-of-lane events in 14 days for this room); premium is restored only by the copy/glob split, not by a one-off pin on this sha.

     Open the signed attestation →

114. 2026-09-02 · 🎭 performer · offPct 18

     feat(cog): /qwen — the offline model, one door (flight lane)

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 13→33 over 6 commits since

     Open the signed attestation →

115. 2026-09-02 · 🎤 voice · offPct 32 · structural

     fix(deck): raise v1.2-r5 — the anti-bounce pass from the operator's own bounce analysis

     Why it left the lane: The diff's real work is entirely rhetorical/persuasion craft — the strings '% of the exposure it unlocks', 'counted is coverable', 'the trigger is a number, not an argument', 'you could have looked — for real', 'We don't bet on model breakthroughs', 'The leave-behind — every number, checkable without us' are exactly the kind of Six-Needs/paradox-voice reframing 🎤 voice does on every commit. But every one of those strings lives inside src/app/deck/raise/RaiseDeckClient.tsx (a React component, JSX + a DECK_VERSION const) and its assertions in tests/deck/raise-deck.test.mjs — neither path appears anywhere in voice's owned-surface list (books/tesseract/chapters/*.md, src/content/blog/*.mdx, docs/05-content/blog/scratchpad/, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/, scripts/voice-*-check.sh). The gzip-NCD reef has almost nothing to match against (prewalkOverlap=2 — near-zero seed overlap before the walk even starts), so the recursive walk had to do essentially all the placement work and landed the mass on the only thing actually present: TSX/JSX and a jest test file, not markdown content.

     The semantic pull: 'DECK_VERSION', 'RaiseDeckClient.tsx', 'assert.match', and the JSX className soup around every edited paragraph are code-shaped tokens with no analog in voice's markdown corpus; they pull the walk toward whatever lane recognizes .tsx + test assertions (builder/architect), while the prose itself (which the reef would happily grip as voice) is buried as JSX children rather than being its own file.

     Intent–reality gap: The message claims four 'surgical' prose fixes — precisely scoped, targeted persuasion-craft edits, which is legitimate voice work — but the receipt measured 32% of 118 lit reality cells outside intent's lane because that legitimate prose work only exists as inline JSX inside a component file voice has never been declared to own.

     • Create src/content/decks/raise-deck-copy.ts with the five extracted copy blocks named in code_rewrite
     • Update RaiseDeckClient.tsx to import from the new module and delete the inline JSX-embedded prose
     • Add src/content/decks/** to voice's owned-surface glob in .workflow/rooms/voice.html
     • Re-run tests/deck/raise-deck.test.mjs against the extracted constants (assertions can match on the imported strings directly instead of scraping rendered flat text)
     • Repeat the extraction for the dentry fork this file mentions (the sibling deck the header notes 'the copy did not leak back into dentry') so both decks share the same content-ownership pattern
     • Add tests/ops/room-surface-match.test.mjs (warn-only) asserting Relevant-Rooms trailers intersect the named room's owned globs for copy-dominant commits

     What we learned: This repo has at least one class of file (investor decks) where voice-craft PROSE and application CODE are physically the same file — room ownership globs assume content and code are separable by path, and wherever a deck stores its copy inline in a .tsx component, that assumption is false and the receipt will keep flagging correct prose work as out-of-lane until the copy is extracted to its own path.

     The missing guard (the deliverable): tests/ops/room-surface-match.test.mjs (warn-only, mirrors rules-for.mjs's advisory posture): for any commit whose message carries a Relevant-Rooms trailer, compute whether the changed file paths intersect that room's declared owned-surface globs; when they don't AND the diff is copy-only (string-literal/JSX-text changes with no control-flow delta), print a suggestion to extract the copy into a content module rather than silently accepting the mismatch.

     Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Right-sized, not shallow or saturated in the truncation sense (100% of walked cells lit on both sides, walkTruncated=false) — but asymmetric: intent's walk terminates at 103 while reality needs 118, suggesting the intent side is under-seeded relative to the richness of the commit message. Concrete adjustment: seed the intent walk from the FULL commit message body (this one is unusually dense — four named fixes, a Story trailer, a Persona-Intent trailer) rather than a truncated head, so intent coverage can grow to match reality's natural spread instead of capping at 103. Experiments: After extracting raise-deck-copy.ts, re-run this same commit's diff through the reef and measure whether prewalkOverlap rises above 2 — if it does, that's a direct before/after receipt that content-extraction fixes reef placement, not just room bookkeeping. · Ablation-test seeding the intent walk from the full commit message (including Story/Persona-Intent trailers) vs. message-head-only on the last 10 voice-room commits; measure whether litIntent moves closer to litReality and offPct drops. · Recompress the last 14 days of 🎤 voice out-of-lane commits (103 events) grouped by whether their changed files are under src/app/** vs the room's declared markdown globs — measure what fraction of the 103 is this exact code-embedded-copy pattern vs a different mechanism, since the checklist above only fixes the deck-specific instance.

     Measured outcome (deterministic, not self-graded): room mean offPct 22→17 over 7 commits since · dictionary ablation saved 1277B

     An out-of-lane copy-craft commit on a live fundraising deck is a false-alarm loss event for this lane, not a real defect — it inflates 🎤 voice's apparent loss ratio for work that was actually correct, and premium is restored the moment the content path is extracted and added to the room's own coverage so the receipt can see the work where it actually is.

     Open the signed attestation →

116. 2026-09-02 · 🎭 performer · offPct 17

     feat(pmu): effective-path-count audit — the certainty denominator priced honestly (participation ratio)

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 12→17 over 6 commits since

     Open the signed attestation →

117. 2026-09-02 · 🎭 performer · offPct 19 · mistagged

     feat(pmu): spectral-radius --sweep — the receipt tracks the resolvent hyperbola to the pole

     Why it left the lane: The diff is 175 lines of pure PMU math/testing vocabulary — SWEEP_GRID, POLE_PROBE_C, resolventTotal, directDiscountedSum, ladderLambda, sweepMeasure, writeSweepSvg, spawnSync-driven firing tests — none of which appears anywhere in 🎭 performer's owned surface (docs/decks/, docs/strategy/deck/, docs/01-business/fundraising/, docs/strategy/munich-re-*, docs/ops/demo-*). The commit message itself is written in dense narrative/metaphor register ('the resolvent hyperbola', 'the pole', 'divergence... made visible', 'the tolerance is measured, not picked') — exactly the storytelling voice CLAUDE.md's pitch/comms rules train into performer's terminal. That register is what seeds the INTENT lane toward performer's narrative coordinates; the REALITY code is 100% Operations/Grid math+testing identifiers that never touch that lane. prewalkOverlap=3 (vs litIntent=134/litReality=109) shows almost no reef-seed agreement going in — the mismatch is real, not noise.

     The semantic pull: Literal terms pulling the wrong way: prose like 'the resolvent hyperbola to the pole' and 'divergence... made visible' (performer/narrative register) vs code identifiers SWEEP_GRID / POLE_PROBE_C / resolventTotal / directDiscountedSum / ladderLambda / spawnSync (pure Operations/PMU-testing register). No file in the diff (scripts/pmu/spectral-radius.mjs, tests/pmu-simulator/spectral-radius-guard.test.mjs) matches any performer-owned glob.

     Intent–reality gap: The commit message and trailers correctly self-declare this as PMU/testing work with 🧪 laboratory and 📐 architect as the PRIMARY Relevant-Rooms (performer is listed third, i.e. Originating-Terminal only) — CLAUDE.md explicitly calls that divergence 'fine and worth recording.' The receipt nonetheless scored it against performer's lane and flagged 19% off, treating an expected, self-declared divergence as a violation.

     • Confirm scripts/pmu/**/*.mjs and tests/pmu-simulator/**/*.test.mjs are present in 🧪 laboratory's or 📐 architect's owned-surface glob in .workflow/rooms/*.html
     • Patch the offPct alarm's room-selection logic to prefer best-glob-match Relevant-Room over Originating-Terminal when Originating-Terminal's own match count is 0
     • Add a regression test: Originating-Terminal glob-match=0 + a Relevant-Room glob-match=100% ⇒ alarm attributes to that Relevant-Room, not Originating-Terminal
     • Re-audit performer's 14 out-of-lane events in the last 14 days for the same signature (PMU/testing-only diff, Originating-Terminal=performer, Relevant-Rooms names laboratory/architect first) — likely the dominant cause of the room's elevated count
     • Verify performer's owned-surface globs genuinely exclude scripts/pmu/ and tests/pmu-simulator/ (no accidental overlap masking this from the other direction)

     What we learned: f9e88143b fixed WHO GETS CREDIT for a cross-terminal commit (read Relevant-Rooms instead of only Originating-Terminal) but not WHETHER THE LANE ALARM FIRES — the offPct threshold check still appears to score against Originating-Terminal's owned surface even when that room's glob-match is zero and a correctly-declared Relevant-Room's glob-match is total. Crediting and gating are two different code paths and only one got the fix.

     The missing guard (the deliverable): A test asserting: for a commit whose Originating-Terminal owns 0% of the changed files but whose first-listed Relevant-Room owns 100% of them, the lane/offPct alarm attributes and scores against that Relevant-Room, not the Originating-Terminal.

     Detector probe: signal attribution — gzip reef 15% · ballistic walk 85%. walkTruncated=true with litReality(109) well below litIntent(134) suggests the reality-side walk is being cut off before it fully expands the cascade for a two-file, math+test+SVG-writer diff that legitimately spans several concepts. Adjustment: re-run this sha with --max-depth increased by 1-2 plies on the reality walk only and compare litReality/offPct before concluding the 19% is semantic drift rather than truncation. Experiments: Re-run this commit's reef seed with PMU vocabulary (SWEEP_GRID, resolventTotal, ladderLambda) added to the shared gzip dictionary and measure the prewalkOverlap delta · Re-run the reality walk on this sha with --max-depth +2 and measure whether litReality grows toward litIntent and offPct drops · Recompute offPct for this commit against laboratory's owned-surface glob instead of performer's and check whether it falls under the 15% threshold — a direct test of the mistagged-room hypothesis

     Measured outcome (deterministic, not self-graded): room mean offPct 12→17 over 6 commits since · dictionary ablation saved 984B

     If this is the room-attribution false-positive class rather than genuine reckless drift, it artificially inflates performer's loss-ratio against well-tested, guard-shipping infra work; premium is restored the moment the lane alarm scores against the glob-matching Relevant-Room instead of the Originating-Terminal.

     Open the signed attestation →

118. 2026-09-02 · 🎤 voice · offPct 48 · structural

     feat(deck): raise v1.2-r4 — manifesto, no-brokers/no-resale, recognition-equals-raise, Hooper moved, sourced

     Why it left the lane: The diff is one dictation pass that stacks four semantically distinct registers inside a single component file. 'ThetaDriven is the attestation pillar...Detected/Placed/Priced/Dispatched' is voice-room prose (matches the book/blog corpus). But 'Mint · $20/agent-yr · anyone, today → Meter · $0.002 per decision → Settlement · paper pays only against the named key' is money-mechanics vocabulary that pulls toward the vault lattice; 'T-28 to T-57', 'December seed (2,000 kinetic = 2-5 accounts)' is spec/sizing arithmetic that pulls toward architect; and the Caremark/T.J.-Hooper legal-precedent move plus the entire parity rewrite in tests/deck/raise-deck.test.mjs is pure architect/guard territory. litReality=142 vs litIntent=134 (8 extra reality cells) and prewalkOverlap of only 4 out of ~134-142 lit cells show the reef seeded reality's mass into different territory from intent's seed almost immediately, before the walk even ran — consistent with vault/architect vocabulary dominating roughly half the lit reality cells.

     The semantic pull: '$0.002 per decision', 'Mint · Meter · Settlement', 'paper pays only against the named key' (fee-stack/money terms) and 'In re Caremark', 'The T.J. Hooper', the whole parity-test rewrite (spec/guard terms) are the two strongest pulls — one toward 🔒 vault's money coordinate, one toward 📐 architect's spec/legal coordinate — while only the 'manifesto'/'recognition' prose blocks are native to 🎤 voice.

     Intent–reality gap: The commit claims a single Originating-Terminal (🎤 voice) for a five-part dictation round that itself lists three Relevant-Rooms (voice, vault, architect); the receipt measured against voice's owned-surface lane alone and found ~48% of the lit reality mass outside it — the commit message already conceded the cross-room shape the lane score then penalized.

     • Add a glob for the extracted frame files to each room's owned-surface config (voice: frames/Manifesto*.tsx; vault: frames/FeeStack*.tsx; architect: frames/Recognition*.tsx, frames/LegalPrecedent*.tsx, tests/deck/**)
     • Until the split lands, mark src/app/deck/raise/RaiseDeckClient.tsx explicitly cross-room (unowned by any single room) rather than silently defaulting into voice's lane
     • Update the commit-trailer convention doc to require a rationale or split when Relevant-Rooms lists ≥3 entries
     • Re-score the room's 14-day backlog (102 events) against the corrected glob to see how much of the 'chronic' offPct was this misconfiguration vs genuine drift
     • Extract the repeated Tailwind boilerplate (see compression_dictionary_adds) into shared className constants so future diffs carry less non-semantic mass

     What we learned: The 🎤 voice room's owned-surface glob (books/blog/decentralized/6needs) has never included any path under src/app/deck/, yet this room's actual recurring work product is dictated raise-deck copy landing in exactly that path — every deck-editing dictation commit is therefore structurally guaranteed to score as out-of-lane regardless of prose quality, which is the more plausible explanation for this room's persistent 20-39% offPct baseline (102 events/14 days) than 102 independent quality failures.

     The missing guard (the deliverable): A pre-receipt check that, for any commit whose message enumerates ≥3 Relevant-Rooms, refuses to score offPct against the single Originating-Terminal's owned-surface glob alone — it must score against the union of the named rooms' globs, or the commit must carry a Lane-Exempt trailer. Guard: tests/ops/cross-room-commit-lane-exempt.test.mjs.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true with litIntent/litReality both in the 130-142 range suggests the walk hit its budget before fully resolving either side's coordinate spread; concrete adjustment: re-run this commit's walk with --max-depth increased one ply and compare whether offPct converges or holds, to separate 'truncated-artifact' drift from real placement drift. Experiments: Re-run the receipt after adding src/app/deck/raise/frames/**-style globs to vault+architect and measure whether offPct for this sha drops toward the ~15 threshold · Ablate the fee-stack/legal vocabulary ('mint', 'settlement', 'Caremark', 'Hooper') from the intent corpus seed and re-walk to confirm it's the dominant pull term, not a correlate · Bump --max-depth by one ply on this sha specifically (walkTruncated=true) and compare litIntent/litReality deltas to see if truncation itself is inflating offPct independent of the room-glob issue

     Measured outcome (deterministic, not self-graded): room mean offPct 22→22 over 4 commits since · dictionary ablation saved 1475B

     A room chronically scoring 20-48% off-lane on an artifact its own owned-surface config never claimed reads as unmeasurable risk rather than genuine drift, inflating the voice room's apparent loss-ratio; restoring the glob (or splitting the file by domain) converts this from a standing false-positive into an honest, scoreable lane.

     Open the signed attestation →

119. 2026-09-01 · 🎭 performer · offPct 26 · structural

     feat(bookclub): render-ticker instrument — @domain #rank ▲x.x%, rank+arrow-only public json

     Why it left the lane: The diff's real strings — scripts/bookclub/ticker.mjs, highscore-publish.sh, src/data/render-ticker.json, tests/bookclub/ticker.test.mjs — share zero vocabulary with 🎭 performer's owned surface (docs/decks/, docs/strategy/deck/, event-*, decentralizedtalk/, fundraising, munich-re-*, over-exposure-*, demo-*). There's no region narrative cached to point at specific stolen coords, which is itself the tell: the reef isn't finding a competing lane to pull toward, it's finding NO lane at all for scripts/bookclub/* and src/data/*, so the walk's reality lit-set (142) spreads past intent (127) into whatever nearest-neighbor coordinates gzip-NCD can match, none of which are performer's.

     The semantic pull: Terms like 'bookclub', 'render', 'ticker', 'domain', 'rank' and the shell idiom in highscore-publish.sh have no counterpart anywhere in the performer glob list — the pull isn't toward a specific wrong lane, it's the absence of any owned lane for this content, so the walk's placement is structurally undetermined rather than semantically misdirected.

     Intent–reality gap: The commit's Originating-Terminal trailer claims 🎭 performer, and the commit message itself is accurate about what it built — but the built artifact (a bookclub render-ticker instrument) has no relationship to performer's owned decks/pitch/event surface at all; the receipt is measuring a room-ownership gap, not a bad claim about the work.

     • Add scripts/bookclub/**, src/data/render-ticker.json, tests/bookclub/** to 🔨 builder's (or ☕ network's, since bookclub sends live there) owned-surface glob in .workflow/rooms/*.html
     • Fix scripts/pmu/intervention-fire.sh:74 (per prior lesson f9e88143b, still unfixed) to attribute against the UNION of Relevant-Rooms, not Originating-Terminal alone — this commit's own Relevant-Rooms trailer already names 🔨 builder and ☕ network correctly
     • Add a pre-commit or nightly check that flags any top-level scripts/<dir>/, src/data/<file>, or tests/<dir>/ path with zero matching glob across all .workflow/rooms/*.html files — new feature directories must register an owner in the same commit that creates them
     • Note in CLAUDE.md's SOVEREIGN OUTREACH ENGINE section (Key scripts table) which room owns scripts/bookclub/* now that it has an owner, so the next session doesn't have to re-derive it

     What we learned: This is a new variant of the ownership gap, distinct from f08ec5875's cross-room orchestration STATE files: an entirely ordinary new PRODUCT feature directory (scripts/bookclub/, src/data/) can be built correctly, tested, and guarded, and still register as out-of-lane for whatever room's terminal happened to author it — because the ownership model has no mechanism that requires a new directory to register a room the moment it's created. The gap isn't in orchestration bookkeeping, it's in ordinary feature growth outpacing the glob list.

     The missing guard (the deliverable): A guard (warn-only, like WIP claims) that diffs the set of top-level paths touched in a commit against the union of every .workflow/rooms/*.html owned-surface glob, and fails/warns when a touched directory matches none of them — forcing the commit that creates a new scripts/<x>/ or src/data/<x> surface to also claim it in some room's glob.

     Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. Likely under-sized for this commit's shape: 4 files across 3 extensions (.sh, .mjs, .json, .mjs-test) touching a brand-new directory gives the walk more surface to place with no anchor, and walkTruncated=true confirms it ran out of budget before settling. Adjustment: for commits creating a new top-level scripts/<dir>/ or src/data/ path (detectable from git diff --diff-filter=A), raise --max-depth or the walk's step budget by one tier so truncation-driven spread doesn't get counted as semantic drift. Experiments: Backfill the owning glob for scripts/bookclub/** in .workflow/rooms/*.html, then re-render this exact sha's receipt and measure whether offPct drops materially below 26 — isolates the ownership-gap contribution from the truncation contribution. · Re-run intervention-fire.sh's attribution logic against Relevant-Rooms-union instead of Originating-Terminal-only across the last 20 out-of-lane events in .thetacog/pipeline-gate-trips.ndjson, and measure how many resolve to in-lane — validates f9e88143b's unfixed prescription with fresh data. · A/B walkTruncated=true vs. a doubled --max-depth for commits with diff-stat >150 lines across >3 new files (like this one), and compare resulting offPct — tests whether truncation or genuine cross-lane content is the dominant driver for this commit shape.

     Measured outcome (deterministic, not self-graded): room mean offPct 17→13 over 12 commits since · dictionary ablation saved 968B

     An out-of-lane event caused by a genuine ownership gap (no room glob covers a correct, guarded, tested feature) is a false loss against this lane's premium — the fix is registering the glob and the attribution logic, not disciplining the room, and once both land this class stops charging performer for work it never actually owns.

     Open the signed attestation →

120. 2026-09-01 · 🎭 performer · offPct 17 · structural

     post(blog): tape-never-says format round — FORMAT GREEN (needs in titles, schematic, sensory dishes, win condition up top)

     Why it left the lane: The diff adds three concrete strings that gzip-compress far better against 🎭 performer's fundraising/demo corpus than against the declared blog/voice/architect lane: (1) the new fenced block rendering `@moonshotscapital.com #2 ▲0.5%` — a receipt-table shape that mirrors the API-output format performer's docs/ops/demo-* scripts print; (2) the literal new line 'Check the melt yourself: `curl -s https://thetadriven.com/api/ticker`' — a live demo command, structurally identical to the curl invocations performer's demo docs carry; (3) the retained/strengthened phrase 'a written path from a twenty-dollar agent-year licence' and repeated 'the milestone board' — verbatim pricing vocabulary that lives in docs/01-business/fundraising/ and docs/decks/insure-rounds/, both on performer's owned surface. None of these three n-grams exist in voice's or architect's corpus, so the reef's gzip-NCD seed pulled reality mass toward performer's coordinate even though the file (src/content/blog/2026-09-01-...) sits entirely outside every declared room's owned-surface glob.

     The semantic pull: 'milestone board', 'agent-year licence', 'twelve-zero path' and the curl-to-/api/ticker demo line are performer-flavored strings embedded in a blog post whose Relevant-Rooms trailer names performer/voice/architect but whose owned-surface globs (docs/decks/, docs/01-business/fundraising/, docs/ops/demo-*) never include src/content/blog/** — so the vocabulary is recognized by the reef but the file location is not, and the two signals disagree.

     Intent–reality gap: The commit claims pure format work ('FORMAT GREEN... this round is structure, not argument') but the receipt shows 17% of lit reality mass landing on coordinates that echo performer's pricing/demo corpus — because the format pass also sharpened a live demo command and tightened pricing phrasing that a pure structural edit wouldn't have touched.

     • Add `Cites-Corpus:` to the COMMIT MESSAGE STRUCTURE trailer list in CLAUDE.md alongside Originating-Terminal/Relevant-Rooms, with the same worked example used here.
     • Extend .workflow/rooms/performer.html (and voice/architect) owned-surface docs with the explicit blog-content exemption clause.
     • Add a citation-exemption branch to the offPct computation and a unit test asserting a commit with a matching Cites-Corpus trailer scores lower offPct than an identical commit without one.
     • Backfill this sha's own commit message with a `Cites-Corpus: 🎭 performer (milestone board pricing, /api/ticker demo)` trailer via `git notes add` (never amend, per the shared-tree rule) so the ledger has one real example to test the guard against.
     • Re-run scripts/pmu/validate-dogfood.sh after the guard lands to confirm no unrelated blog commit is now silently exempted for real scope creep.

     What we learned: Prior lessons fixed trailer-reading (f9e88143b) and named ownerless shared-state files (f08ec5875) — both are FILE-location gaps. This event is a different structural gap even with a correct trailer and a correctly-owned-vs-unowned file split: the reef's gzip-NCD placement is purely LEXICAL, so a blog post that legitimately quotes another room's pricing/demo vocabulary (not scope creep, just citation) reads as drift with no mechanism to declare 'this overlap is intentional and bounded.' The room model conflates file-location ownership with vocabulary-based reef placement and has no citation escape valve.

     The missing guard (the deliverable): tests/ops/citation-exemption.test.mjs — asserts a commit carrying a Cites-Corpus trailer naming room R excludes R's cited terms from its own offPct tally, and asserts a commit with the SAME vocabulary but no trailer still counts as off-lane (so the exemption can't be abused as a blanket suppressor).

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized, not shallow or saturated — litIntent/litReality track each other closely (134 vs 132) so the walk isn't starving or flooding. The fix belongs upstream of the walk: feed the seed corpus a wider intent basin (via the Cites-Corpus exemption) rather than deepening max-depth or widening the walk itself. Experiments: Re-render sha 309fc74d7's receipt after synthetically adding a Cites-Corpus trailer for performer's milestone-board/demo vocabulary; confirm offPct drops below 15 with the exemption logic applied. · Add 'milestone board', 'agent-year licence', 'the maître d\', presenting:' to the compression dictionary and recompress the last 20 commits' receipts with/without them; measure whether reef-wide mean offPct shifts (ablation test, per the dictionary-adds protocol). · Run intervention-fire.sh's citation-exemption branch against 5 known-good (correctly in-lane) and 5 known-bad (genuinely out-of-lane) historical commits to confirm the exemption doesn't silently launder real drift.

     Measured outcome (deterministic, not self-graded): room mean offPct 14→12 over 6 commits since · dictionary ablation saved 1015B

     A structural false-positive like this inflates 🎭 performer's measured loss-ratio on content it never actually touched, which is restored the moment the citation-exemption guard ships and this room's off-lane count is re-baselined against declared, not incidental, vocabulary overlap.

     Open the signed attestation →

121. 2026-09-01 · 📐 architect · offPct 20 · mistagged

     feat(iamfim): pay off the steering wheel on the left door

     Why it left the lane: The entire 32-line diff lands in src/app/iamfim-landing/page.tsx — landing-page prose ("steering wheel", "left door"/"right door", "D&O register", the tests/surfaces/iamfim-determinism-is-reproducibility.test.mjs guard) — none of which touches architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, CLAUDE.md). The commit's own trailers already name the mismatch: Originating-Terminal is 📐 architect but Relevant-Rooms lists 🎤 voice as a co-claimant, and the reef's prewalkOverlap=6 (low relative to litIntent=129) shows the seed was placed against architect's strategy-doc lane before the walk ever ran, so the walk inherited a wrong starting lane rather than discovering one.

     The semantic pull: Terms like "steering wheel", "left door", "weather", "felt vs analytical" are voice/copy vocabulary that gzip-NCD would seed toward the voice/builder cell, not architect's strategy-doc cell — but Originating-Terminal alone pins the intent lane to architect before that pull can register.

     Intent–reality gap: The commit message and trailers correctly describe a landing-page copy edit and even list 🎤 voice in Relevant-Rooms, but the receipt's lane classification appears to key off Originating-Terminal (📐 architect) alone — so work that CLAUDE.md's own commit-trailer protocol declares as legitimately divergent gets scored as drift.

     • Ship scripts/pmu/room-drift-<room>-check.sh (or equivalent) reading Relevant-Rooms as an OR-set, not just Originating-Terminal
     • Add tests/pmu-simulator/room-lane-relevant-rooms.test.mjs asserting a commit with Relevant-Rooms containing the reality-lane room is NOT flagged out-of-lane
     • Update .workflow/rooms/architect.html 'Owned file/dir surface' section with an explicit note: Relevant-Rooms divergence is expected-normal per CLAUDE.md, not drift
     • Add a 'lane-source' field to the receipt row (originating vs relevant-room match) so future intervention passes can see which classifier path fired
     • Audit whether the ff09774e5 spec actually landed as a merged commit — if not, this event is proof it hasn't, and that itself is the thing to fix first

     What we learned: This is the third independent spec of the identical fix (Originating-Terminal-only lane matching vs CLAUDE.md's sanctioned Relevant-Rooms divergence) with zero shipped code across all three — the room mean offPct trend (14→28→26→20, no monotonic improvement) shows that a spec graduating into prose/CLAUDE.md-adjacent docs is not the same as a landed commit against the actual detector, which is exactly the DELEGATE-TO-SUBTASKS rule's own caveat about a coordinate staying open until a commit lands against it — applied here to the detector's own maintenance debt.

     The missing guard (the deliverable): A guard that fails if no committed script/test in this repo implements Relevant-Rooms-aware lane matching for the drift receipt — i.e., a guard on the fix's own existence, since three prior specs described it and none shipped code.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow: every walked cell lit on both sides (129/129, 127/127) and walkTruncated=true, meaning the walk hit its depth/time budget before it could spread wide enough to separate a mistagged seed from genuine drift. Adjustment: raise --max-depth (or the time budget) specifically for commits carrying multiple Relevant-Rooms trailers, since those are the cases most likely to need extra spread to disambiguate lanes. Experiments: Re-run this commit's reef seed using Relevant-Rooms-weighted seeding instead of Originating-Terminal-only and measure whether offPct drops below 15 for this and similar mixed-trailer commits · Raise walk --max-depth for commits with >1 Relevant-Rooms entry and measure whether walkTruncated=false while litIntent/litReality convergence improves · Ablation-test seeding the trailer template strings (Originating-Terminal:, Relevant-Rooms:) into the gzip dictionary and measure whether prewalkOverlap rises for correctly-mixed-lane commits without also rising for genuinely out-of-lane ones

     Measured outcome (deterministic, not self-graded): room mean offPct 16→21 over 4 commits since · dictionary ablation saved 1482B

     A mistagged-not-actually-drifted commit like this inflates architect's apparent loss ratio for free; premium only resets once the lane detector reads Relevant-Rooms and the fix ships as a merged guard rather than a fourth unshipped spec.

     Open the signed attestation →

122. 2026-09-01 · 🎭 performer · offPct 20 · mistagged

     feat(track): self-device cookie — the operator's own phone stops paging him

     Why it left the lane: The diff touches exactly three files — src/app/api/activity/track/route.ts, src/app/api/track/visit-shape/route.ts, tests/surfaces/self-device-excluded.test.mjs — and every load-bearing string in it ("request.cookies.get('td_self')", "NextResponse.json", "shouldSendEmail", "self_device: true", the ordering-assertion test file itself) is backend API / test vocabulary. None of it touches performer's owned globs (docs/decks/, docs/strategy/deck/, docs/strategy/event-*, decentralizedtalk/, docs/01-business/fundraising/, munich-re-*, over-exposure-*, docs/ops/demo-*). The commit's own trailer already says as much: Relevant-Rooms is 🔨 builder, 📐 architect, 🎩 operator — performer is not even listed. The reef's gzip-NCD placement pulled the mass toward builder/architect coordinates because that is literally what the diff is; Originating-Terminal: 🎭 performer only reflects which app the operator happened to be typing in.

     The semantic pull: route.ts / cookies.get('td_self') / NextResponse.json / event_data.self_device / tests/surfaces/*.test.mjs — pure API-route-and-guard vocabulary, zero deck/pitch/fundraising/event terms anywhere in the diff. There is no region narrative cached for this sha, so the specific out-of-lane coords aren't available; the strongest evidence is the room's own owned-surface glob list, against which 0 of 3 changed files match.

     Intent–reality gap: The commit message frames this as a clean engineering fix and is filed under the performer room via Originating-Terminal, but the receipt measured reality landing entirely inside builder/architect's territory — a divergence the commit's own Relevant-Rooms trailer already disclosed and CLAUDE.md explicitly calls 'fine and worth recording,' yet the lane pipeline scored it as 20% out-of-lane drift anyway.

     • Add src/app/api/activity/track/** and src/app/api/track/** to builder's (or architect's) owned-surface glob in .workflow/rooms/*.html
     • Add tests/surfaces/** to whichever room legitimately owns cross-cutting API test guards
     • Patch the lane-scoring/receipt pipeline to parse the Relevant-Rooms trailer and either re-lane against it or exempt commits where Originating-Terminal is explicitly outside it
     • Re-run the receipt for 53dce715e scored against Relevant-Rooms=builder/architect/operator to confirm offPct drops under 15
     • Cross-reference this incident in the COMMIT MESSAGE STRUCTURE section of CLAUDE.md, next to the existing note that Originating-Terminal and Relevant-Rooms 'can diverge and that's fine'

     What we learned: CLAUDE.md already sanctions Originating-Terminal ≠ Relevant-Rooms divergence as normal and useful (the operator can commit builder/architect work while physically sitting in the performer terminal, and the trailer is supposed to make that legible) — but the trailer is convention, not a gate, and the receipt/lane pipeline has no mechanism to read it. So the pipeline is penalizing the exact, correctly-self-disclosed behavior CLAUDE.md designed the trailer to accommodate. This is a different failure than f08ec5875's ownerless-shared-state gap: this file has clear, legitimate owners (builder/architect); the gap is that the lane model can't consume the one field that already resolves the ambiguity.

     The missing guard (the deliverable): A guard/step in the receipt pipeline that parses the commit's Relevant-Rooms trailer and, when it excludes Originating-Terminal, either re-scores offPct against Relevant-Rooms or emits a distinct 'self-declared cross-room' verdict instead of an out-of-lane drift flag for the terminal room.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized on saturation (fully lit, litIntent==walksIntent and litReality==walksReality) but walkTruncated=true with reality outrunning intent by 12 cells (136 vs 124) suggests the reality-side walk hit the depth/time budget slightly later than the intent side. Concrete adjustment: bump --max-depth (or the ply time budget) symmetrically for both sides on this class of commit so truncation lands at matched depth, and re-check whether the 12-cell asymmetry itself was manufacturing part of the apparent drift. Experiments: Re-run this receipt with lane=Relevant-Rooms (builder/architect/operator) instead of lane=Originating-Terminal (performer) and measure whether offPct falls under the 15 threshold — validates the mistagged-lane hypothesis directly. · Add builder/architect owned-surface globs for src/app/api/activity/track/** and src/app/api/track/**, then re-render this same commit's panel and confirm offPct against performer specifically becomes moot (the commit no longer needs to be graded against performer's lane at all). · Increase --max-depth for the reality-side walk by one ply and re-measure litReality vs litIntent parity; if the 12-cell gap collapses, truncation asymmetry (not seed placement) was a real contributor and the 65/35 split above should shift toward walk.

     Measured outcome (deterministic, not self-graded): room mean offPct 20→35 over 4 commits since · dictionary ablation saved 728B

     This event's loss is entirely attribution noise, not defective work — the fix (glob additions + trailer-aware lane scoring) restores performer's premium to baseline without touching a single line of the shipped tracking fix, which stays correct and guarded as-is.

     Open the signed attestation →

123. 2026-09-01 · 📐 VS Code Architect · offPct 41 · mistagged

     post(audit-invariant): the invariant IS trigger-shaped, and the repo IS the attestation

     Why it left the lane: Both changed paths — docs/05-content/blog/cook-rounds/2026-09-01-the-audit-invariant.claims.md and src/content/blog/2026-09-01-the-audit-invariant.mdx — sit entirely outside Architect's declared owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The prose that pulled the reef match is strategy-flavored vocabulary living in a blog post: 'the audit invariant is trigger-shaped', 'a policy whose TRIGGER is settled by argument', 'rating variable', 'insurability', 'premium/loss-ratio' — these are exactly the terms that seed docs/strategy/ and docs/01-business/strategy/ corpora, so the reef found a lexical match to Architect's lane even though the physical files are 100% blog content. The commit's own Relevant-Rooms trailer already ranks Architect third of three (🎤 Voice, 🎩 Operator, 📐 Architect) — the commit's own metadata knew this before the receipt did.

     The semantic pull: Strategy/underwriting register terms ('trigger-shaped', 'rating variable', 'measured, binary, cause-blind, and settled without adjudication', 'the repo IS the attestation') read like docs/strategy content and dragged the reef seed toward Architect's owned vocabulary, while the actual diff hunks are entirely under src/content/blog and docs/05-content/blog/cook-rounds — paths Architect's globs never claim.

     Intent–reality gap: The commit's Originating-Terminal trailer claims Architect-lane work, but the receipt measured 100% of the lit reality mass (offPct=41, well past the 15 threshold) landing in files no Architect glob covers — this is a blog editorial correction, not strategy-doc work, wearing the wrong room's trailer because that's the terminal the operator happened to be typing in.

     • Patch the room-lane scorer to prefer Relevant-Rooms[0] over Originating-Terminal whenever Originating-Terminal's owned-surface has zero overlap with changed files.
     • Add the sentence above to CLAUDE.md COMMIT MESSAGE STRUCTURE clarifying which trailer field the receipt reads.
     • Audit the other 52 Architect out-of-lane events in the last 14 days — if most are docs/05-content/blog or src/content/blog paths, this single fix retires the bulk of the room's flagged debt.
     • Confirm/add an explicit owned-surface glob for docs/05-content/blog/** and src/content/blog/** on whichever room (likely 🎤 Voice, already ranked first in this commit's Relevant-Rooms) actually owns content work, in `.workflow/rooms/*.html`.
     • Add the missing_guard test in the same commit as the scorer fix.

     What we learned: The receipt attributes room-lane by Originating-Terminal (physical location) while the commit already computes a better answer in Relevant-Rooms (content-domain match) in the same trailer block — CLAUDE.md explicitly sanctions the two diverging, but nothing tells the receipt pipeline which one is authoritative for lane-scoring, so a purely physical-desk signal is silently driving a content-drift metric. This is a new failure class distinct from the prior three facts-only lessons: it isn't the work drifting, it's the ATTRIBUTION FIELD being wrong.

     The missing guard (the deliverable): A test asserting: for any commit where changed-file globs have zero overlap with Originating-Terminal's owned surface AND nonzero overlap with a Relevant-Rooms entry ranked above it, the room-lane offPct must be computed against that higher-ranked Relevant-Rooms entry, not Originating-Terminal — catching the terminal-vs-content-domain mismatch mechanically instead of letting it accumulate as false out-of-lane debt.

     Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. Right-sized on the reality side but ceiling-bound on intent — walksIntent equals litIntent at the full 144-cell cap, meaning Architect's intent corpus is broad enough to saturate the lattice on every check, masking how narrow its true owned-surface (6 globs) actually is. Adjustment: tighten Architect's intent-walk seed/budget so it reflects the 6-glob narrowness instead of saturating to the ceiling, which would make offPct on genuinely narrow lanes more discriminating. Experiments: Recompute this commit's offPct against a content-domain pseudo-lane (docs/05-content/blog/** + src/content/blog/** as the corpus) instead of Architect's corpus, and confirm offPct collapses toward 0 — validates the mistagged-not-defective hypothesis directly. · Re-key all 53 of Architect's last-14-day out-of-lane events by Relevant-Rooms[0] instead of Originating-Terminal and measure how many disappear — tests whether the routing fix explains the bulk of the room's flagged debt. · Track prewalkOverlap specifically for Architect commits split by content-vs-strategy file type over the next 30 days; if content commits cluster at low prewalkOverlap and strategy commits cluster high, the reef can auto-flag terminal-vs-content mismatches from seed overlap alone before the walk even runs.

     Measured outcome (deterministic, not self-graded): room mean offPct —→6 over 6 commits since · dictionary ablation saved 933B

     A mistagged event like this reads as loss history against Architect's lane even though the room did nothing wrong, inflating its loss ratio on a false trigger; premium is restored the moment lane-scoring reads Relevant-Rooms[0] instead of Originating-Terminal for content-domain paths, so Architect's offPct trend reflects genuine strategy-doc drift rather than terminal-location noise.

     Open the signed attestation →

124. 2026-09-01 · 📐 VS Code Architect · offPct 34

     book(ch12): the floor and the ceiling — the halt is real, and it is not what is sold

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 24→8 over 5 commits since

     Open the signed attestation →

125. 2026-09-01 · 📐 architect · offPct 37 · mistagged

     fix(signup): blog signups notify the operator — the intake route was silent

     Why it left the lane: Every changed line sits in src/app/api/book-club/subscribe/route.ts (the new `import { notifyOperator }` and the `await notifyOperator(...)` call) and tests/api/operator-notify-coverage.test.mjs (the new COVERED array entry) — zero lines touch docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, or CLAUDE.md, which is architect's entire owned surface. The reef seed (prewalkOverlap=4 against litIntent=128, ~3%) found almost no overlap between those terms and whatever corpus represents the architect strategy-doc lane, so the walk propagated from a nearly-unseeded intent side while reality lit 115 cells cleanly around 'notify/route/coverage' vocabulary — the classic signature of a commit whose Originating-Terminal trailer names a room the diff never visits.

     The semantic pull: Terms 'notifyOperator', 'route.ts', 'COVERED', 'operator-notify-coverage.test.mjs' are pure application/test vocabulary; they gzip-compress toward code-lattice cells the reef associates with implementation rooms, not with architect's docs/strategy cells — there is no architect-owned string in the diff for the walk to anchor on.

     Intent–reality gap: The commit trailers claim this is 📐architect-lane work (Originating-Terminal + Relevant-Rooms both list architect), but the receipt measured 100% of the changed mass landing in src/app/api and tests/api — files no architect glob covers; the work itself is correct, the room attribution on it is not.

     • Ship the glob-based `room-drift-<room>-check.mjs` that three rooms (architect, navigator, laboratory per ff09774e5) each independently spec'd but none shipped — this event is the fourth occurrence of the same gap
     • Wire that script into the same receipt pipeline that produced this offPct=37 row, so lane classification reads owned-surface globs instead of trusting the trailer alone
     • Follow up on the commit's own flagged debt: replace the hand-typed COVERED array in tests/api/operator-notify-coverage.test.mjs with a computed scan of client fetch()->route.ts pairs (the commit measured 40/52 pairs don't notify and explicitly declined to fix the enumeration)
     • Add an explicit note to whichever room's owned-surface file legitimately covers src/app/api/**/route.ts and tests/api/** so gap-closing commits like this one classify correctly going forward
     • Cross-link this event's sha to ff09774e5 in the lesson ledger so the next re-derivation of the room-drift-check finds prior art instead of re-speccing from scratch a fourth time

     What we learned: The drift detector's own room-classification is itself an instance of THE PROXY IS NOT THE THING pattern this very commit fixed: it uses the hand-typed Originating-Terminal trailer as a stand-in for 'which room's owned surface was touched,' exactly parallel to the hand-typed COVERED array the commit message calls out as 'a proxy for the construct, not the construct.' The fix (compute lane from file globs) has been independently re-derived three times without shipping — this commit is fresh, self-aware evidence that the same failure mode now sits one layer up, inside the tooling that measures drift.

     The missing guard (the deliverable): A test asserting that for every commit, at least one room named in Relevant-Rooms has an owned-surface glob (`.workflow/rooms/*.html`) matching at least one changed file path — a computed check, not an enumerated one, mirroring `tests/ops/rules-for.test.mjs`'s own construct-vs-token discipline.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized, not shallow or saturated: litIntent=128 vs litReality=115 is only a ~10% asymmetry with no truncation flag. No depth adjustment needed here; the defect is upstream in what seeds the walk, not in how far it travels. Experiments: Re-run prewalkOverlap for this sha seeding the intent corpus from a file-glob-matched room (whichever owns src/app/api) instead of the Originating-Terminal-only corpus, and measure whether overlap rises materially above 4 · Backtest the proposed glob-based lane classifier against the room's other 191 out-of-lane events from the last 14 days to measure what fraction would reclassify as in-lane · Ablation-test adding this commit's notify/coverage vocabulary to whichever room's dictionary actually owns src/app/api, and confirm offPct drops on a re-render of this sha under that assignment

     Measured outcome (deterministic, not self-graded): room mean offPct 25→20 over 3 commits since · dictionary ablation saved 884B

     A correct, well-reasoned bug fix registering as 37% out-of-lane inflates this room's apparent loss ratio with false-positive drift rather than real risk; premium is restored not by changing the code but by shipping the glob-based lane check so the detector stops charging architect for work it did correctly, just outside its nominal doc surface.

     Open the signed attestation →

126. 2026-09-01 · 🎭 performer · offPct 18 · structural

     post(blog): tape-never-says round 1 — specialist accuracy fixes + three sub-bar courses revised

     Why it left the lane: Every changed line in this commit sits inside src/content/blog/2026-09-01-the-tape-never-says-what-it-measures.mdx — a file that matches NONE of performer's owned globs (docs/decks/, docs/decks/insure-rounds/, docs/strategy/deck/, docs/01-business/fundraising/, docs/strategy/event-*/, docs/strategy/munich-re-*, docs/strategy/over-exposure-*, docs/ops/demo-*, decentralizedtalk/, .claude/commands/deck.md). The prose itself carries dense pricing/finance vocabulary — 'transaction price vs published-methodology indices', 'S&P/VIX', 'a twenty-dollar agent-year licence to a number with twelve zeroes', 'share of a market, not a reading off a gauge' — which is exactly the register performer's fundraising/insure-rounds/deck corpus is built from. The reef's gzip-NCD seed reads that lexical family and drags the reality walk toward performer's pricing-deck coordinates even though the physical file lives entirely outside performer's surface.

     The semantic pull: Finance/pricing nouns in the diff ('transaction price', 'S&P and the VIX disclose their methodology', 'licence', 'twelve zeroes', 'share of a market') gzip-match performer's deck/insure-rounds/fundraising corpus far more tightly than they match any blog-content corpus, because no room's owned-surface glob claims src/content/blog/** at all.

     Intent–reality gap: The commit message claims a pure content-accuracy edit to a published blog post (specialist audit fixes, ghost-read reception fixes) with Relevant-Rooms correctly naming performer/voice/architect — but the receipt measures 18% of lit reality mass landing outside ANY declared lane, because none of those three rooms' owned-surface manifests actually include blog content as a claimed glob.

     • Add `src/content/blog/**` glob to whichever `.workflow/rooms/*.html` is the voice room's manifest
     • Add `books/tesseract/**` to the same or a book-specific room if book content is currently also ownerless
     • Re-verify f9e88143b's fix (room-activity.mjs / intervention-fire.sh reading Relevant-Rooms) actually ran for this sha — offPct=18 suggests either it wasn't applied here or all three Relevant-Rooms still lack the glob
     • Add a structural guard asserting every top-level content genre directory resolves to at least one owning room
     • Re-run this commit's lane computation after the glob patch and record the new offPct as the fix's receipt

     What we learned: This is a third distinct flavor of the ownerless-file gap (beyond f9e88143b's wrong-trailer bug and f08ec5875's orchestration-state gap): PUBLISHED CONTENT genres (src/content/blog/**, books/tesseract/**) have zero owning-room glob anywhere in .workflow/rooms/*.html, so any terminal editing blog/book prose scores maximal off-lane purely from an ownership blind spot — independent of whether Relevant-Rooms is populated correctly.

     The missing guard (the deliverable): A test that walks .workflow/rooms/*.html and asserts every top-level content-bearing directory (src/content/blog/, books/tesseract/, docs/comms/) is claimed by at least one room's owned-surface glob — failing loudly when a content genre has zero claimants.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent ≈ 1.0 and litReality/walksReality ≈ 1.0 (both sides fully lit relative to walks taken), yet walkTruncated=true — the walk is hitting its depth cap while still fully lighting every walked cell, which looks like early saturation rather than natural convergence. Concrete adjustment: raise --max-depth for src/content/blog/** commits specifically and check whether offPct drops once the walk is allowed to run past the current cap. Experiments: Recompress the last 20 blog-post commits with a content-corpus separation (blog prose tokens weighted apart from deck/pricing tokens) and measure whether offPct falls once finance vocabulary in blog posts stops gzip-matching performer's fundraising corpus · Re-run this commit with --max-depth raised past the current truncation point and record whether offPct converges lower once the walk isn't capped · Gate lane computation on prewalkOverlap/litIntent ratio: for commits below ~25% pre-walk overlap (like this one, 27/139≈19%), fall back to a Relevant-Rooms weighted-average lane before trusting the reef-seeded placement, and measure recurrence over the next 10 room events

     Measured outcome (deterministic, not self-graded): room mean offPct 14→12 over 6 commits since · dictionary ablation saved 773B

     An 18%-off-lane content commit with no owning glob inflates performer's apparent loss ratio for work it never actually did; restoring the premium just requires the voice-room glob patch plus a re-scored receipt showing the same commit landing in-lane once ownership exists.

     Open the signed attestation →

127. 2026-09-01 · 📐 VS Code Architect · offPct 17 · mistagged

     cook(format): completeness is a WHOLE-DOCUMENT score, and C24 is only its local half

     Why it left the lane: The diff touches exactly two files — docs/05-content/blog/mdx-cook-format.md (a check's rule-spec doc) and scripts/blog/mdx-cook-check.sh (the check script itself) — and neither string in the diff is architecture vocabulary. The load-bearing terms are 'COMPLETENESS', 'the LOCAL half', 'sufficiency', 'DOMAIN SPECIALIST', 'does the argument CLOSE' and the six-score say-block in mdx-cook-check.sh — all content-voice-craft and check-tooling vocabulary. The room's own file (.workflow/rooms/vscode-architect.html:622) already names the identical class — 'scripts/voice-*-check.sh — owned by 🔨 Builder (shipped script) and 🧪 Lab Cursor (rule spec)' — as FORBIDDEN surface for architect. mdx-cook-check.sh is structurally the same object (a check script + its rule-spec doc), and terminal-voice.html:582-589 confirms docs/05-content/blog/* and scripts/*-check.sh rule-spec ownership sits with Voice/Lab/Builder, not architect. The walk lit heavily into that vocabulary (litReality=108, fully saturated) because that IS where the bytes are.

     The semantic pull: 'COMPLETENESS', 'sufficiency', 'DOMAIN SPECIALIST', 'does the argument CLOSE', and the six-score say() block in mdx-cook-check.sh — pure content-format/voice-craft terminology with zero overlap to architect's owned vocabulary (docs/strategy/, docs/03-product/, CLAUDE.md cardinal-goal language).

     Intent–reality gap: The commit trailers themselves already disagree with the Originating-Terminal tag: Relevant-Rooms lists 🧪 Cursor Laboratory and 🎤 Terminal Voice ahead of 📐 VS Code Architect, and the room's own forbidden-surface list (vscode-architect.html:622) already excludes exactly this file-class (check-script + rule-spec pairs). The commit message claims nothing about architecture at all — it is a pure content-format correction — yet Originating-Terminal:architect is what the receipt used as the claimed lane, so the receipt is scoring physical terminal against content territory the room's own doc says isn't its lane.

     • Generalize the Forbidden-surface bullet in .workflow/rooms/vscode-architect.html from the single named 'scripts/voice-*-check.sh' example to the class 'scripts/**/*-check.sh + paired docs/**/*-format.md'
     • Implement the spec'd-but-missing scripts/room-drift-architect-check.sh named at vscode-architect.html:635 — it has been spec'd, unbuilt, since before this event
     • Key that hook off (Originating-Terminal, top Relevant-Room) mismatch, not physical-terminal alone, so commits where Relevant-Rooms already names the correct owner are caught before the receipt has to
     • Audit the room's other 51 out-of-lane events in the last 14 days for the same check-script/rule-spec-doc signature — this may be one recurring class rather than 52 distinct ones
     • Add mdx-cook-check.sh and mdx-cook-format.md explicitly to cursor-laboratory.html's and terminal-voice.html's owned-surface lists (currently only scripts/voice-*-check.sh is named, not scripts/blog/*)

     What we learned: The room's own architecture doc already anticipates and forbids this exact class (scripts/voice-*-check.sh at vscode-architect.html:622) but the forbidden-surface list is enumerated by literal path rather than by CLASS, so a new sibling check-script (mdx-cook-check.sh) isn't covered until someone names it too — the rule was written narrow enough to be dodged by the next file with the same shape.

     The missing guard (the deliverable): scripts/room-drift-architect-check.sh — already spec'd at vscode-architect.html:635 ('flags any staged file outside the owned surface when the commit message includes room: architect or the session is open in VS Code') but marked 'not yet implemented.' Build it, and key it off BOTH Originating-Terminal and the commit's own Relevant-Rooms trailer, since this event's trailer already named the correct owner (Laboratory) and a terminal-only check would have caught it just as well.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent=100% and litReality/walksReality=100% — every cell the walk reached, it lit; nothing died out naturally. Combined with walkTruncated=true, the walk is under-provisioned for this commit: it hit its depth/time ceiling before it could find whatever deeper overlap exists, rather than converging on its own. Adjustment: raise --max-depth for commits touching long append-only doc files like docs/05-content/blog/mdx-cook-format.md and re-measure whether litIntent/litReality converge before truncation. Experiments: Re-run this commit's walk with --max-depth +2 and diff the resulting offPct against the shipped 17 to test whether truncation, not seed mismatch, is the larger driver · Score lane-fit against the commit's own Relevant-Rooms trailer (Laboratory here) in parallel with Originating-Terminal for the room's last 52 events, and measure whether the trailer-based score reduces the flagged count · Seed the reef corpus with this diff's repeated phrases ('does the argument CLOSE', 'DOMAIN SPECIALIST') tagged to Laboratory/Voice territory, then re-measure prewalkOverlap on the next mdx-cook-format.md commit to see if it rises above 2

     Measured outcome (deterministic, not self-graded): room mean offPct 20→41 over 3 commits since · dictionary ablation saved 1017B

     A mistagged, low-severity crossing (offPct 17 vs threshold 15, on a room already running 52 events/14 days) doesn't move the loss-ratio much on its own, but the recurring count means the lane's premium stays inflated until the forbidden-surface rule is generalized to the check-script CLASS and the already-spec'd drift hook is actually shipped — until then every new sibling check-script re-trips the same claim.

     Open the signed attestation →

128. 2026-09-01 · 📐 VS Code Architect · offPct 37 · structural

     cook(format): the 95% bar is FOUR scores, and two of them are new

     Why it left the lane: The new C27 block in scripts/blog/mdx-cook-check.sh is a banned-claim regression scanner, and its literal regex bank quotes fragments from at least six OTHER CLAUDE.md hard-rule domains inside one file: 'reads directly from the CPU hardware performance counter' (AXIOM 0 apparatus-scope), 'guarantee ... ai|agent|model|system' (NEVER GUARANTEE BEHAVIOR), 'safe harbou?r', 'keeps? ... in lane', 'airtight|bulletproof defense' (legal-outcome), 'software cannot audit software' (banned voice construction). A single commit whose own text restates the vocabulary of six unrelated lattice cells is guaranteed to light reality-side coordinates far outside the 'mdx-cook format' lane it claims, because the reef seed (gzip-NCD) sees those strings as belonging to their origin rules, not to this one.

     The semantic pull: Literal strings pulling mass elsewhere: '(reads?|read) (directly )?from the (cpu|hardware) (performance )?(counter|register)', 'guarantee|guarantees|guaranteed) (that )?(the |your )?(ai|agent|model|system)', 'safe harbou?r', 'keeps? (it|them|the agent|your agents?) in ?(the )?lane', 'airtight|bullet ?proof (defen[cs]e|defense)' — each is a near-verbatim quote of another hard rule's own banned-phrase language, so the reality-walk necessarily fans across those rules' coordinates rather than concentrating in the content-format lane the commit message describes.

     Intent–reality gap: The message claims scoped work — 'the 95% bar is FOUR scores' for the mdx-cook grading loop — but the reality lane the receipt measured is a multi-domain regression scanner whose entire job is to detect regressions of OTHER rules, so its reality footprint is structurally wider than the single-lane intent the commit describes.

     • Add the multi-domain-scanner note to docs/05-content/blog/mdx-cook-format.md's C27 section
     • Extract C27's regex bank into scripts/voice/banned-claims-bank.mjs and re-point tests/voice/no-behavior-guarantee.test.mjs + tests/pmu-simulator/axiom0-grounding.test.mjs to import it
     • Assign scripts/blog/** and docs/05-content/blog/** to the room whose owned-surface actually covers content tooling (Relevant-Rooms already names 🎤 Terminal Voice and 🧪 Cursor Laboratory — neither currently lists these globs either; audit and fix)
     • Add tests/ops/no-duplicate-banned-claim-banks.test.mjs guarding against a second independent regex bank for an already-guarded claim
     • Clarify in CLAUDE.md's COMMIT MESSAGE STRUCTURE section whether offPct/lane grading should key off Originating-Terminal or Relevant-Rooms when they diverge (the section already permits divergence but the receipt pipeline appears to still grade against Originating-Terminal only)
     • Re-run mdx-cook-check.sh's receipt with the C27 cross-domain strings temporarily stripped to confirm they are the offPct driver before committing to the extraction

     What we learned: A file whose PURPOSE is to scan for regressions of other rules (a ledger, an index, a banned-claim scanner) will always show elevated offPct under this pipeline, because its content legitimately quotes many other lattice cells — the receipt currently has no way to distinguish 'this commit drifted' from 'this commit is a correctly-designed cross-cutting scanner,' and for this room specifically, the receipt appears to grade against Originating-Terminal ownership even though CLAUDE.md's own commit-trailer convention explicitly sanctions Originating-Terminal and Relevant-Rooms diverging.

     The missing guard (the deliverable): tests/ops/no-duplicate-banned-claim-banks.test.mjs — scans shell/JS/mjs files for banned-claim regex literals labelled the same as an existing tests/voice/*.test.mjs pattern (by shared substring, e.g. 'safe harbour', 'guarantee', 'apparatus-scope') and fails if two files independently hard-code near-duplicate patterns for the same claim instead of importing one shared module.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent ≈ 1.0 and litReality/walksReality ≈ 1.0 — essentially every walk step lit a new cell with no revisits/saturation on either side, meaning the walk is expanding cleanly but was not run deep enough to reach a convergence or plateau point. Concrete adjustment: bump --max-depth by one ply on a re-run of this sha and check whether litIntent/litReality converge (shared vocabulary catches up) or diverge further (confirming the cross-domain-scatter read). Experiments: Re-render this commit's receipt with the C27 cross-domain regex strings (hardware/guarantee/safe-harbour/keep-in-lane phrases) stripped from the reality corpus and measure whether prewalkOverlap rises and offPct drops — isolates the vocabulary-scatter hypothesis directly. · Tag files whose role is 'regression scanner over other rules' (a recognizable class: mdx-cook-check.sh C27, any future banned-claim bank) with a header the ingest step can detect, and score their reality lane as a UNION of the domains they scan rather than a single claimed lane — ablation-test against this sha and 3 similar scanner commits. · Re-attribute this room's last 49 out-of-lane events against Relevant-Rooms instead of Originating-Terminal and compare the resulting offPct distribution — if the mean drops materially, the terminal-vs-relevance conflation is the dominant driver for this room, not genuine drift.

     Measured outcome (deterministic, not self-graded): room mean offPct 29→18 over 7 commits since · dictionary ablation saved 1103B

     An event this explainable — a correctly-designed cross-domain scanner registering as drift because the pipeline grades against the wrong room field — is a measurement-calibration loss, not a code-quality loss, and this room's premium is restored not by tightening the code further but by fixing which field (Originating-Terminal vs Relevant-Rooms) the offPct is scored against.

     Open the signed attestation →

129. 2026-09-01 · 📐 VS Code Architect · offPct 25

     book(ch12): The Second Theorem — the DPI half, which the book argued nowhere

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 26→22 over 9 commits since

     Open the signed attestation →

130. 2026-09-01 · 📐 VS Code Architect · offPct 44 · structural

     copy(iamfim): the kicker gets a second beat — the invoice hands them the wheel

     Why it left the lane: The diff is 65 lines of pure copywriting rationale on src/app/iamfim-landing/page.tsx — strings like 'kicker', 'parenthetical', 'steering wheel', 'normal-case tracking', 'invoice comes with a steering wheel' — landing-page/marketing-copy vocabulary with no strategy-doc counterpart in this room's owned surface. The commit message that carries the INTENT signal is, paradoxically, dense with this room's own strategic register (Rice, 'claim perimeter', 'DETECTED, PLACED, PRICED, DISPATCHED', 'AMNESIA LEDGER register', 'banned sentence') because that's simply how this operator writes rationale — so the walk seeds intent heavily into docs/strategy-adjacent coords while the actual changed file sits in product/marketing-copy territory the Architect glob never covers. prewalkOverlap=5 against ~139-140 lit cells (~3.6%) shows the two seeds started in almost disjoint regions before the recursive walk even ran.

     The semantic pull: Strategic essay-register terms ('claim perimeter', 'Rice', 'the banned sentence', 'AMNESIA LEDGER') anchor litIntent inside Architect's declared lane (docs/strategy/); the touched path src/app/iamfim-landing/page.tsx and copy-craft terms ('kicker', 'parenthetical', 'steering wheel') anchor litReality outside it — two vocabularies sharing one commit, pulling in different lattice directions.

     Intent–reality gap: The commit is scoped copy(iamfim) — a landing-page marketing-copy edit — but is attributed to 📐 VS Code Architect, whose owned surface is docs/strategy/, docs/03-product/, .workflow/rooms/*.html and CLAUDE.md; the receipt correctly measures that the actual file (src/app/iamfim-landing/page.tsx) sits outside that declared lane, even though the commit's own Relevant-Rooms trailer already names 🎭 Alacritty Performer as a better content match.

     • grep scripts/pmu/ for where Originating-Terminal is parsed into a lane glob and confirm Relevant-Rooms is currently ignored there
     • extend that parser to union Relevant-Rooms' owned-surface globs into the intent lane
     • add the CLAUDE.md sentence above pinning Relevant-Rooms as the lane source, not Originating-Terminal
     • audit the room's last 47 out-of-lane events and bucket how many carry a copy(/scope-prefixed) commit type vs true strategy-doc drift, to size the fix's expected impact
     • re-run this sha's receipt with a higher --max-depth to check whether walkTruncated=true is masking a lower true offPct before treating 44 as the final number

     What we learned: The commit-message convention already has a field built for exactly this divergence (Relevant-Rooms), and CLAUDE.md itself says the divergence is 'fine and worth recording' — but the drift-receipt pipeline never reads that field; it silently computes lane purity from Originating-Terminal alone. The tooling and the doc disagree with each other, and nobody noticed because the prior two interventions on this room ran facts-only (no sensemaking).

     The missing guard (the deliverable): A test asserting that for a commit carrying a Relevant-Rooms trailer, the computed intent lane equals the union of ALL named rooms' owned-surface globs, not just Originating-Terminal's — feed it a synthetic commit (Originating-Terminal=Architect, Relevant-Rooms includes Performer, file under Performer's glob) and assert offPct drops once Relevant-Rooms is consulted.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true means both sides hit the depth/time budget ceiling near-simultaneously at ~139-140 cells — the walk is SATURATED, not organically converged, so it's unclear whether offPct=44 would settle lower (more overlap found deeper) or hold if allowed to keep walking. Adjustment: re-run this sha with --max-depth raised past the current ceiling and compare; if offPct converges toward 15 the number was a truncation artifact, if it holds near 44 the drift is real. Experiments: Re-run sha d545b7681e with --max-depth raised until walkTruncated=false and record whether offPct moves — settles the saturation-vs-real question directly · Modify lane assignment to union Relevant-Rooms globs for the last 90 days of Architect-room commits and re-measure the room's offPct distribution — test whether the 47-event/14-day rate collapses once Relevant-Rooms is consulted · Compare prewalkOverlap across all `copy(` scope-prefixed commits vs all docs/strategy-prefixed commits from this room to check whether low prewalkOverlap (~5) is a consistent fingerprint of copy-type commits specifically, which would justify pre-classifying them before the walk runs

     Measured outcome (deterministic, not self-graded): room mean offPct 44→17 over 4 commits since · dictionary ablation saved 500B

     This room's 47 out-of-lane events in 14 days look like elevated loss frequency but are largely a pipeline-attribution artifact (lane computed from Originating-Terminal, ignoring the commit's own Relevant-Rooms declaration); fixing the lane source restores an accurate premium signal by removing false-positive drift from the count rather than requiring any change to the underlying copywriting work.

     Open the signed attestation →

131. 2026-09-01 · 🎭 performer · offPct 27 · mistagged

     fix(ticker): flat arrow prints 0.0%, never a dash

     Why it left the lane: The diff's real vocabulary — TrendMark, '▲{n}%', '▼{Math.abs(n)}%', '0.0%', 'flat arrow', 'held exactly steady', 'tape' — is pure market-signal-renderer language. It gzip-matches hardest against B3.Tactics.Signal (pull ×11, the heaviest single coordinate) and spreads into C1/C2/C3 Operations.Grid/Loop/Flow (the render-loop/tape mechanics) and B1/B2 Speed/Deal (movement, exchange). None of that vocabulary shares any mass with performer's owned surface (docs/decks/, docs/strategy/, event scripts, fundraising docs) — the commit was typed in the Alacritty performer terminal but its content has zero lexical overlap with performer's domain.

     The semantic pull: TrendMark/▲/▼/0.0%/tape are signal-and-flow terms with no analog in performer's deck/capital vocabulary, so the reef seed (prewalkOverlap=2 — almost no pre-walk agreement between intent and reality) placed reality far from the performer-claimed intent lane before the ballistic walk even ran.

     Intent–reality gap: The commit message and trailers correctly scope this as a one-line UI fix and even declare Relevant-Rooms: performer, builder, architect — an honest multi-room claim. The receipt nonetheless measures it against performer's lane alone (offPct=27) because the firing script room-resolves off Originating-Terminal only, discarding the very Relevant-Rooms trailer CLAUDE.md mandates as the content-relevance signal.

     • scripts/interventions/intervention-fire.sh:74 — replace the standalone sed extraction with a call into (or a port of) room-activity.mjs's dual-trailer resolver so the two implementations cannot diverge again
     • ​.workflow/rooms/alacritty-performer.html — broaden the Forbidden Surface bullet from the single named file to `src/components/**` generally
     • Verify iterm2-builder.html's Owned file/dir surface already claims `src/components/**` explicitly; add it if the glob is only implicit
     • Implement scripts/room-drift-performer-check.sh — named in alacritty-performer.html:644 as 'spec'd, not yet implemented' and is exactly the guard this event needed
     • tests/ops/intervention-fire.test.js — add a fixture: Originating-Terminal performer + Relevant-Rooms performer,builder + a src/components/*.tsx diff, asserting the lane check runs against builder's surface, not performer's

     What we learned: The prior fix for this exact mechanism (f9e88143b) landed in scripts/rooms/room-activity.mjs's dual-trailer resolver but was never wired into scripts/interventions/intervention-fire.sh, which still runs its own one-line Originating-Terminal-only sed at line 74. The room-resolution logic has FORKED into two implementations — one correct, one not — which is exactly the 'one rule lives in one place' failure CLAUDE.md itself names, applied to the pipeline's own room-attribution code rather than to prose.

     The missing guard (the deliverable): A consistency test asserting intervention-fire.sh's ROOM= resolution equals room-activity.mjs's resolveRoom() output for a fixture commit set with divergent Originating-Terminal vs Relevant-Rooms trailers — so the two implementations cannot silently re-diverge.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. walkTruncated=true on a 4-line, 1-file diff is oversized relative to the change: a near-zero-LOC commit hitting the depth/time budget cap suggests walk depth isn't scaled to diff size, letting a tiny commit's walk wander as far as a large one's. Adjustment: scale --max-depth (or the ply budget) to changed-LOC so single-line/few-line commits get a shallower, more contained walk. Experiments: Re-run intervention-fire's room resolver against the last 20 out-of-lane performer events with the Relevant-Rooms-aware logic and measure whether offPct would have stayed under 15 for each — a retroactive ablation of this exact fix · Tag UI-ticker vocabulary (TrendMark, ▲/▼, %, tape) in the reef dictionary explicitly under Builder/Architect lanes so future ticker-adjacent commits seed correctly regardless of which terminal authored them · Measure prewalkOverlap across all 9 performer out-of-lane events in the last 14 days — if it's consistently ≤3, that confirms reef-seed (not walk-spread) is the dominant driver for this room specifically, prioritizing the seed-level fix over walk-depth tuning

     Measured outcome (deterministic, not self-graded): room mean offPct 16→18 over 3 commits since · dictionary ablation saved 532B

     A one-line, correctly-scoped bugfix reading as a 27%-off-lane loss event inflates performer's measured breach rate with false signal, diluting the premium calculation's trust in the room's actual risk profile until the resolver reads Relevant-Rooms and this class of misattribution stops contributing noise to the loss ratio.

     Open the signed attestation →

132. 2026-09-01 · 📐 VS Code Architect · offPct 29 · structural

     cook(format): the cooking IS the memory palace — plus axes 5 and 6, and C27 narrowed

     Why it left the lane: Every string that actually shipped in this diff is content-craft/voice-craft vocabulary, not strategy vocabulary: `cook(format):` in the subject, the touched paths `.claude/skills/mdx-cook/SKILL.md`, `docs/05-content/blog/mdx-cook-format.md`, `scripts/blog/mdx-cook-check.sh`, and body text like 'the cooking IS the memory palace', 'the courses are ROOMS', 'C24 is the completeness check', 'C27 narrowed', 'the specialist read is the fact gate's discovery mechanism'. None of those three touched paths intersect any glob in the room's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) — a 0-for-3 path match. No region narrative is cached for this sha, so I can't cite the exact [row,col] anchors the walk lit, but the diff-vs-owned-glob mismatch is itself positional evidence: the reef had nothing in the Architect lane to pull this content toward, so the walk's mass had nowhere to land but content/voice-craft coordinates — which the receipt then scores as 'outside the claimed lane' at offPct=29.

     The semantic pull: The commit's own Relevant-Rooms trailer already names the true owners in rank order — 🎤 Terminal Voice, 🧪 Cursor Laboratory — with 📐 VS Code Architect listed third, weakest. The pull toward those two rooms' coordinates is the diff's entire vocabulary (memory-palace/cooking/six-scores/specialist-read craft rules); nothing in the diff pulls toward Strategy or CLAUDE.md.

     Intent–reality gap: The commit's Originating-Terminal trailer claims the 📐 Architect lane (strategy/business-doc surface), but the receipt measured reality landing in content-format/voice-craft coordinates that match zero of Architect's owned globs — while the same commit's own Relevant-Rooms trailer had already, correctly, named Voice and Laboratory as the true owners.

     • Add the Relevant-Rooms-overrides-Originating-Terminal rule to CLAUDE.md's COMMIT MESSAGE STRUCTURE section (spec_rewrite above)
     • Locate and patch the room-resolver script that currently keys the receipt's lane on Originating-Terminal alone
     • Confirm/add mdx-cook + blog-format globs (.claude/skills/mdx-cook/**, docs/05-content/blog/mdx-cook-format.md, scripts/blog/mdx-cook-check.sh) to the owned-surface declaration in whichever of .workflow/rooms/*.html represents Terminal Voice or Cursor Laboratory
     • Add a guard test asserting: diff with 0% overlap on Originating-Terminal's owned glob + 100% overlap on a Relevant-Rooms glob resolves the lane to that Relevant-Rooms entry
     • Re-render this sha's receipt against the corrected lane and confirm offPct drops toward the intent-lane floor

     What we learned: CLAUDE.md's own trailer convention deliberately decouples Originating-Terminal (where the operator sat) from Relevant-Rooms (who owns the content) and calls that divergence 'fine and worth recording' — but the drift receipt's lane check apparently still keys offPct off Originating-Terminal alone, so any content-craft commit authored from a non-owning terminal is guaranteed to read as out-of-lane regardless of how well-scoped the work actually is. The measurement disagrees with the architecture's own stated design.

     The missing guard (the deliverable): A test that constructs a commit with Originating-Terminal X (0% owned-glob overlap with the diff) and Relevant-Rooms naming Y (100% owned-glob overlap), and asserts the receipt's lane/offPct computation resolves against Y, not X.

     Detector probe: signal attribution — gzip reef 55% · ballistic walk 45%. Both sides are fully lit relative to their own walk budget (100% lit-to-walked ratio) with no truncation — this is a saturation signature, not an under-exploration one. Given docsOnly=false was set for a diff that is overwhelmingly two markdown files plus a 21-line shell-script edit, the classifier may be over-weighting the diff as code-reef material and granting it more walk depth/reach than a near-docs-only commit needs. Concrete adjustment: audit the docsOnly threshold against this diff's actual code fraction and re-run with docsOnly forced true to see if litReality contracts toward litIntent. Experiments: Re-render sha e76db302f with the lane check scored against Relevant-Rooms' owned globs instead of Originating-Terminal's, and measure whether offPct falls from 29 toward the intent-lane floor · Audit why docsOnly=false fired on a 2-markdown+1-shell-script diff; re-render with docsOnly forced true and compare litReality/offPct against the current run · Sample prewalkOverlap across the last 20 commits from this room to establish whether 2 is anomalously low for content-craft-from-Architect commits specifically, versus this room's baseline — if systematically low, argues for a dedicated Voice/Laboratory reef rather than walk-depth tuning

     Measured outcome (deterministic, not self-graded): room mean offPct 24→8 over 5 commits since · dictionary ablation saved 946B

     An out-of-lane receipt on well-scoped, correctly-attributed content work (per the commit's own Relevant-Rooms trailer) is a false positive that inflates this lane's loss-ratio for free — restoring it means fixing the scoring field, not the work, which is the cheapest possible premium recovery available here.

     Open the signed attestation →

133. 2026-09-01 · 🎭 performer · offPct 35 · structural

     feat(track): visit-shape persists the shape it mails — scroll depth becomes a series

     Why it left the lane: The lit reality cells are pulled entirely by backend/data-model vocabulary from the diff — `event_data:`, `dwell_ms`, `max_scroll_pct`, `self_device: true`, the `route.ts` insert call, the `tests/surfaces/self-device-excluded.test.mjs` assertions — none of which shares any term with 🎭 performer's owned surface (docs/decks/, docs/strategy/deck/, decentralizedtalk/, docs/01-business/fundraising/, docs/strategy/event-*, docs/ops/demo-*). The commit's own `Relevant-Rooms: 🔨 builder, 📐 architect, 🎩 operator` trailer already names where this content actually belongs — but `scripts/interventions/intervention-fire.sh:74` derives ROOM solely from `Originating-Terminal`, so the offPct/lane gate scored this commit's telemetry-persistence reality against performer's deck/pitch lane, the one lane it was never claiming to occupy.

     The semantic pull: `event_data:`, `dwell_ms`, `max_scroll_pct`, `lanes`, `self_device: true` in the diff, plus the file paths `src/app/api/track/visit-shape/route.ts` and `tests/surfaces/self-device-excluded.test.mjs` — pure Operations/backend vocabulary with zero overlap against performer's pitch-deck/strategy owned-surface globs.

     Intent–reality gap: The commit message and its own Relevant-Rooms trailer both correctly claim this is builder/architect/operator telemetry work (persisting maxScrollPct/dwellMs/lanes to event_data so a scroll-depth series can exist); the receipt nonetheless measured it against 🎭 performer's owned lane purely because Originating-Terminal named performer — a lane the diff shares zero owned globs with.

     • scripts/interventions/intervention-fire.sh: extract Relevant-Rooms trailer alongside Originating-Terminal (line 74) and pass both through
     • scripts/interventions/build-intervention-prompt.mjs: buildPrompt() should accept relevantRooms[] and score/report offPct against the union of those rooms' owned surfaces, not just the room arg it's called with
     • scripts/interventions/probe-verify.mjs:34-38 has the same Originating-Terminal-only room lookup for its 'same-room mean offPct' comparison — fix in the same pass or its before/after trend will keep mixing mistagged and genuine events
     • reuse room-activity.mjs's roomSource/roomReason pattern (trailer vs owned-surface, preserved not collapsed) instead of re-deriving room logic from scratch in intervention-fire.sh
     • CLAUDE.md COMMIT MESSAGE STRUCTURE: append the Relevant-Rooms-is-authoritative-for-scoring clarification (see spec_rewrite)
     • add the regression guard named in missing_guard, in the same commit as the intervention-fire.sh fix

     What we learned: The pipeline's own room-attribution script (intervention-fire.sh:74) ignores the Relevant-Rooms trailer that CLAUDE.md itself mandates and explicitly documents as the content-relevance signal (room-activity.mjs already built and comments this exact split). So any commit correctly authored cross-terminal — which CLAUDE.md calls 'fine and worth recording' — will ALWAYS false-positive as out-of-lane against the terminal it was typed in, regardless of code quality. This is a new gap from the prior lessons (which were about ownerless shared bookkeeping files): here the ownership answer exists, is correctly declared on the commit, and the detector simply reads the wrong trailer.

     The missing guard (the deliverable): tests/interventions/relevant-rooms-scored-not-originating-terminal.test.mjs — SMOKE: construct a synthetic commit whose Originating-Terminal is room A, whose Relevant-Rooms names room B, and whose diff touches only files under room B's owned-surface globs; run intervention-fire's lane/offPct classification and assert it does NOT fire an out-of-lane event against room A. NO-SMOKE: a commit whose diff genuinely touches neither A's nor B's owned surface must still fire, so the guard proves the fix discriminates rather than just silencing the gate.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Fully utilized on both sides — litIntent/walksIntent and litReality/walksReality are both 1.0, meaning the walk saturated every reachable cell rather than dying early or starving. walkTruncated=true means it still hit a depth/frame ceiling somewhere. Concrete adjustment: re-run this sha with --max-depth doubled and diff litReality's extra 8 cells (117-109) against litIntent — if they collapse toward shared territory the offPct was a truncation artifact; if they stabilize at genuinely different coordinates it confirms this is a real content-lane mismatch (consistent with the room-misattribution finding above), not a walk-budget problem. Experiments: Re-score this sha and its 6 sibling out-of-lane events from the last 14d (the room's repeat-offender count) against Relevant-Rooms' owned-surface union instead of Originating-Terminal's, using the existing receipt rows (no re-render needed) — measure whether room mean offPct drops now that mistagged cross-terminal commits are excluded from the performer trend. · Seed the reef's intent-corpus ingest with the Relevant-Rooms trailer text in addition to the commit body, and measure whether prewalkOverlap rises above 4 for commits like this one where the trailer already names the correct content-domain rooms. · Re-run this sha's walk at 2x --max-depth and measure whether the 8-cell litReality/litIntent delta converges (truncation artifact) or holds stable (genuine cross-lane content) — settles the walk_utilization question empirically rather than by inference.

     Measured outcome (deterministic, not self-graded): room mean offPct 30→11 over 8 commits since · dictionary ablation saved 752B

     This false-positive class inflates 🎭 performer's measured loss-ratio on work it never actually owns while masking the room's true drift rate on commits genuinely authored there; premium is restored the moment the scorer reads Relevant-Rooms instead of Originating-Terminal, since only genuine content-lane divergence should carry a cost.

     Open the signed attestation →

134. 2026-08-31 · 🎤 voice · offPct 20 · structural

     feat(milestones): add "Form investable vehicle" — the pitch track had nowhere for a cheque to land

     Why it left the lane: The diff is entirely data (data/milestones-*.json), no semantic-file text available to the receipt, so the ballistic walk has nothing but the room's history and structural JSON shape to anchor to. litReality=132 vastly outpaces litIntent=115 with prewalkOverlap only 3 — the walk found almost no seed overlap between what the commit message claims (adding one milestone predicate, 'Form investable vehicle') and what the JSON diff actually contains: a 1102-line rewrite of milestones-market-contact.json and a 7163-line rewrite of milestones-progress.json. Those two files carry contact/market and progress-tracking vocabulary (deal, allocator, signal, exchange, velocity, throughput) that pulls hard into B1/B2/B3 (Tactics.Speed/Deal/Signal) and C1/C2/C3 (Operations.Grid/Loop/Flow) — 31 encircled cells, weighted toward B3.Tactics.Signal (×11) and C1.Operations.Grid (×6) — while the commit message only narrates a single milestone-board edit (git-discipline / C,B3 lane).

     The semantic pull: 'milestones-progress.json' (7163-line diff) and 'milestones-market-contact.json' (1102-line diff) are bulk regenerated/reordered artifacts, not the 34-line targeted add the message describes (data/milestones.json +34 is the only diff that matches the claim); the other two files' sheer mass drags the walk into Tactics.Deal/Signal and Operations.Grid/Loop/Flow territory that the one-sentence milestone-add message never mentions.

     Intent–reality gap: The commit claims a single, scoped intent — add one milestone predicate ('Form investable vehicle') to the board — but the receipt measures reality landing across two large derived-progress/contact files that dwarf the actual milestone-definition diff by two orders of magnitude (7163+1102 changed lines vs 34).

     • Add a header comment (or a sibling .meta.json) atop milestones-progress.json and milestones-market-contact.json stating '// GENERATED — do not hand-edit; regenerated by <script> whenever data/milestones.json changes' so both the receipt tooling and future readers can classify their diffs as derived, not authored
     • In the generator script (find via `grep -rl milestones-progress.json scripts/`), emit a one-line stderr/stdout summary of what changed (e.g. 'progress: 12 milestones re-scored') so commit messages can cite it mechanically instead of omitting it
     • Extend the commit-message template for milestones-board edits to include a 'Derived files touched:' trailer line listing regenerated JSON outputs separately from the hand-authored delta
     • Register data/milestones-progress.json + data/milestones-market-contact.json as generated files in the NEVER WRITE STATE INTO A GENERATED FILE registry search (`grep -l writeFileSync.*milestones-progress scripts/`) so this room's convention matches the rest of the repo

     What we learned: A commit whose message describes a small, precise intent (34-line add) but whose diff is dominated by regenerated derived-data files (7163 + 1102 lines) will always read as massive out-of-lane drift to the receipt, because the receipt has no signal that distinguishes 'authored delta' from 'mechanical recompute triggered by the delta' — the derived-artifact problem (already named in CLAUDE.md's NEVER WRITE STATE INTO A GENERATED FILE rule) also breaks the drift receipt itself, not just data integrity.

     The missing guard (the deliverable): A pre-commit or receipt-time check that, for any commit touching both an authored source file (data/milestones.json) and its known-generated dependents (data/milestones-progress.json, data/milestones-market-contact.json), requires the commit message to carry a 'Derived files touched:' trailer — absence flags the commit for manual receipt review instead of letting it silently register as room drift.

     Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. Right-sized in ply count (2 plies matches the time budget) but starved on seed quality — with prewalkOverlap=3 the walk is bootstrapping off almost nothing shared between intent and reality text, so its spread (132 lit reality cells) reflects diffuse structural JSON content rather than a strong semantic anchor. Concrete adjustment: for docsOnly=true / data-only commits, increase the seed weight given to the commit message body (Story: trailer) relative to raw diff content, since the diff here is JSON with little natural-language signal for gzip-NCD to grip. Experiments: Recompute this same commit's receipt with the 'Derived files touched:' trailer proposed above added to the message and measure whether offPct drops below 15 — a direct ablation of the spec fix · Run the receipt pipeline over the last 10 milestones-board commits filtering diffs to just data/milestones.json (excluding the two regenerated files) and compare offPct distribution against the unfiltered version, to quantify how much of this room's 93-event-in-14-days out-of-lane rate is this exact derived-file-mass artifact · Add a docsOnly-style flag 'generatedFileMass' computed as (bytes changed in known-generated paths / total bytes changed) to the receipt row, and check whether offPct correlates with it across the last 120 commits (the n120 series already sampled)

     Measured outcome (deterministic, not self-graded): room mean offPct 16→18 over 4 commits since

     This lane's loss ratio takes a mechanical, false-positive hit — a scoped 34-line intent reads as a 20%-off-lane, 31-region breach purely because two regenerated files ship alongside it undeclared; premium is restored not by tightening the room's actual behavior but by making the receipt able to discount derived-file recompute as expected churn.

     Open the signed attestation →

135. 2026-08-31 · 🔨 builder · offPct 19 · mistagged

     style(iamfim): equal doors, and the border carries the qualifier

     Why it left the lane: The commit message declares itself "style(iamfim)" — a narrow CSS/layout lane — but the diff itself introduces THREE distinct changes bundled together: (1) a genuine layout/CSS fix (min-h-[2.8em] on both rail <p> tags, relative positioning), (2) net-new persuasive/legal microcopy ("if you'd rather read the code", "you know who you are" — audience-sorting language the commit message itself analyzes at length as CONDITION vs RECOGNITION register), and (3) a copy deletion ("the same licence, the same checkout"). Items 2 and 3 are content/comms-lane work wearing a style-lane label. That mismatch between the claimed narrow lane and the actual copy-authoring content is exactly what pulls litReality's 132 lit cells 19% outside the 143-cell intent footprint.

     The semantic pull: Real strings from the diff doing the pulling: the two new <span> literals ('if you'd rather read the code', 'you know who you are') and the 20-line JSX comment block above them that explicitly reasons about audience psychology ('CONDITION... because a developer self-selects', 'RECOGNITION... because nobody in that seat needs the role explained') — this is copywriting/positioning prose, not styling prose, sitting inside a commit typed 'style'. No region narrative is cached for this sha, so the exact encircled coords aren't nameable, but the content class (persuasive/legal audience-copy vs CSS) is visible directly in the diff.

     Intent–reality gap: The commit message claims a pure visual fix ('style(iamfim): equal doors, and the border carries the qualifier') scoped to layout, but the receipt measured reality landing 19% outside that lane — because the actual diff also authored new marketing/legal copy and removed an existing line of pricing copy, which is a different kind of work than the claimed CSS fix.

     • Split this class of commit going forward: CSS/layout changes and copy/microcopy changes get separate commits even when discovered in the same review pass.
     • Extract BorderQualifierLabel component (copy) out of the layout JSX in src/app/iamfim-landing/page.tsx.
     • Add a CLAUDE.md line under COMMIT MESSAGE STRUCTURE: a `style(...)`-typed commit may not introduce new user-facing string literals beyond a trivial threshold — use `content(...)` or split.
     • Re-run tests/surfaces/iamfim-fork-ends-at-checkout.test.mjs (already referenced in the file's own comment) to confirm the label spans don't break the checkout-fork invariant.
     • Check whether other 🔨 builder commits this cycle bundle 'style' fixes with new copy the same way — this may be a repeated pattern, not a one-off.

     What we learned: The commit-type prefix (style/fix/feat) is itself an intent-lane claim the pipeline should be able to check against the diff's content shape — a 'style' commit that adds new persuasive string literals is claiming a narrower lane than the work it actually contains, and prior facts-only interventions in this room (offPct trending 16→46, 15→16) never named this mechanism because they had no sensemaking pass to catch it.

     The missing guard (the deliverable): A pre-commit lexical check: when the commit subject's conventional-commit type is `style` or `fix`, flag (warn, not block) if the diff adds new JSX string-literal text nodes (i.e. new user-facing copy) beyond a small character threshold — forcing either a `content`/`feat` type or a commit split.

     Detector probe: signal attribution — gzip reef 25% · ballistic walk 75%. litIntent=143 out of a 144-cell lattice is functionally saturated — the intent side lit almost every possible anchor, which is what happens when a commit message this rich (Story + Persona-Intent trailers, a long diagnostic narrative) gets ingested whole. A saturated intent grid has no headroom to discriminate 'in lane' from 'out', which is consistent with 19% of reality still reading as out-of-lane even against an almost-everywhere intent footprint. Adjustment: trim trailer prose (Story:/Persona-Intent:) out of intent ingest for this commit class, or lower --max-depth when the message body exceeds a length threshold, so intent stays discriminating rather than near-total. Experiments: Re-render 0b4956b547 with the Story:/Persona-Intent: trailers stripped from intent ingest and compare litIntent/offPct against the shipped 143/19 baseline. · Re-render with the 20-line JSX rationale comment excluded from reality ingest (treat comments as docs-lane, not code-lane) and measure whether litReality/offPct shifts — tests whether long in-code prose is itself dragging reality into the same lane the intent trailers occupy. · Pull the last 30 days of commits by conventional-commit type (style/fix/feat/content) and compare mean offPct per type — tests the mistagged-commit-type hypothesis at population scale rather than on this one event.

     Measured outcome (deterministic, not self-graded): room mean offPct 21→11 over 3 commits since · dictionary ablation saved 1172B

     A 19%-off-lane 'style' commit that actually ships new legal/audience-targeting copy is a small, recoverable loss-ratio event on its own, but the room's 62 out-of-lane events in 14 days is the real premium risk — the guard above converts the mistagged-type pattern from invisible to caught-at-commit, which is what restores the room's rate.

     Open the signed attestation →

136. 2026-08-31 · 📐 architect · offPct 17 · mistagged

     feat(funnel): measured verdict — length was the defect, ownDoor 0%→100%, and 95 is above the ceiling

     Why it left the lane: The diff touches exactly two files — docs/funnel/hiring-funnel-ratchet-v1.md and scripts/funnel/aggregate.mjs — and neither path matches any glob in 📐architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md). The pull terms are the persona-probe vocabulary itself — 'ownDoor', 'wouldAct', 'refused', 'inverted', 'salesperson/underwriter/hacker', the regex rewrite in SHARED.wouldAct — none of which are Strategy-axis language; they're operational instrumentation for a live A/B harness. The commit's own Story trailer names the coordinate outright: 'C1.Operations.Grid' — the author already placed this work in Operations, not Strategy, and the reef/walk correctly followed that semantic gravity into the encircled out-of-lane region while the room trailers still claimed architect.

     The semantic pull: Regex/field-name vocabulary ('wouldAct','ownDoor','refused','inverted') plus the file path 'scripts/funnel/aggregate.mjs' and the explicit self-declared coordinate 'C1.Operations.Grid' in the Story line all cluster toward Operations/experiment-instrumentation, dragging litReality away from any of architect's Strategy-prefixed globs.

     Intent–reality gap: The commit claims architect (Strategy) authorship via its Originating-Terminal/Relevant-Rooms trailers, but its own Story sentence names the coordinate 'C1.Operations.Grid', and the actual changed files (docs/funnel/, scripts/funnel/) sit in neither architect's, operator's, nor navigator's declared owned surface — the trailer and the content disagree with each other before the receipt ever measures anything.

     • Add docs/funnel/** and scripts/funnel/** as owned surface to .workflow/rooms/operator.html (or whichever room's data/rooms.json coordinate field maps to Operations) — currently unclaimed by any room
     • Ship the room-drift-<room>-check.sh fix spec'd three times (architect, navigator, laboratory per lesson ff09774e5) and still not unified — this event is a second live instance of the same unshipped gap
     • Add a Story-coordinate-vs-Originating-Terminal cross-check to hooks/post-commit or hooks/pre-commit, reading data/rooms.json's per-room coordinate field against the Story trailer's leading ShortLex prefix
     • Audit prior docs/funnel/*.md and scripts/funnel/*.mjs commits for the same architect-mistag pattern now that the surface gap is named
     • Decide whether 'funnel' experimentation deserves its own owned-surface entry rather than living ambiguously between architect/operator

     What we learned: The self-correcting data already exists inside every commit and is unused: the Story trailer states the ShortLex coordinate axis in plain text ('C1.Operations.Grid') in the SAME commit whose Originating-Terminal/Relevant-Rooms trailers contradict it — the detector doesn't need an external room-to-path mapping to catch this class, it needs to parse the Story line's own coordinate prefix against data/rooms.json's per-room coordinate field, which the prior three unshipped room-drift-check specs never did.

     The missing guard (the deliverable): A commit-message linter (pre-commit or a `tests/comms/story-coordinate-matches-room.test.mjs`) that extracts the leading ShortLex prefix from the Story trailer (e.g. 'C1.Operations.Grid') and the Originating-Terminal room's `coordinate` field from data/rooms.json, and warns (or blocks) when they diverge — turning the self-admitted mismatch visible in this commit's own prose into a mechanical check instead of a receipt discovered after the fact.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litReality/walksReality = 100% (122/122) — every walked cell lit, fully saturated, yet walkTruncated=true means the walk hit its depth/time budget before covering as much ground as intent side (134 walks). For a mixed doc+code commit like this (65 md lines + 2 code lines spanning two unrelated directories), the walk is too shallow: raise --max-depth or the time budget specifically for commits whose diff spans >1 top-level directory. Experiments: Seed docs/funnel/** and scripts/funnel/** as an explicit reef anchor cluster, re-run this commit's receipt, and measure whether prewalkOverlap rises above the current 4 · Seed the Story trailer's literal coordinate string ('C1.Operations.Grid') into the reef as a high-weight ground-truth anchor for commits carrying it, and measure whether litReality shifts toward matching that stated axis directly · Re-run this commit with --max-depth raised one level (walkTruncated=true today) and measure whether litReality closes the 134-vs-122 gap with intent

     Measured outcome (deterministic, not self-graded): room mean offPct 22→32 over 4 commits since · dictionary ablation saved 500B

     17% off-lane mass under a room (architect) whose owned surface never claimed docs/funnel/ or scripts/funnel/ inflates that lane's apparent loss-ratio for risk it was never underwriting; restoring it means the funnel glob moves to the room the Story trailer already names (Operations) so the premium is priced against the party actually holding the work.

     Open the signed attestation →

137. 2026-08-31 · 🔨 builder · offPct 22 · structural

     fix(iamfim): they thought it meant obedient — take the word away in one line

     Why it left the lane: The diff splits almost evenly between src/app/iamfim-landing/page.tsx (builder's owned lane) and tests/surfaces/iamfim-determinism-is-reproducibility.test.mjs (28 vs 24 lines) — a path outside every glob in 🔨 builder's declared owned-surface list. The reality-side reef ingest reads identifier/comment terms from the test file — 'reproducibilityClaimAt', 'resultIdentity', 'elsewhereOrLater', 'gone red on a REWORDING three times' — which project onto a testing/QA coordinate, not the copy/UI coordinate the commit claims. prewalkOverlap=2 (out of 137 walked cells) shows the two corpora barely shared a seed before the walk even ran, which is consistent with one lane being prose-copy and the other being regex/detector engineering.

     The semantic pull: Terms like 'const resultIdentity = /\bsame\s+...', 'RESULT-IDENTITY', 'ELSEWHERE-OR-LATER', and the block comment 'THIS DETECTOR HAS GONE RED ON A REWORDING THREE TIMES' are detector-calibration vocabulary, structurally distinct from the page-copy vocabulary ('obedient', 'grip', 'the car does the exact same thing on ice') that the commit message and lane claim are actually about — the reef has no shared coordinate for 'we are tuning the thing that watches the copy' and 'we are writing the copy.'

     Intent–reality gap: The message claims a single copy-diagnosis fix ('take the word away in one line') but the receipt measured a commit that is nearly 50% guard-engineering prose (three paragraphs of regex-calibration rationale) — legitimate, necessary work, but a second concern the intent language never named as part of the claimed lane.

     • Add `tests/surfaces/*.test.mjs` to .workflow/rooms/builder.html's owned-surface glob so paired guard edits stop registering as cross-lane
     • Audit the other 8 rooms' .workflow/rooms/*.html owned-surface lists for the same gap — any room whose src/app/**/page.tsx carries a 'Guarded by tests/surfaces/...' comment but whose glob excludes tests/surfaces/ will hit this same class
     • Add a naming-convention note near the top of tests/surfaces/ (or in CLAUDE.md's FILE MANAGEMENT section) stating that a guard test's owning room is the room that owns the file it guards, not a separate 'testing' room
     • Consider whether the reef ingest should ingest 'Guarded by tests/...' comment lines from src/app files as a bridge signal linking the two paths at seed time, ahead of the walk

     What we learned: A paired guard-test file (tests/surfaces/*.test.mjs, 1:1 with one page under src/app/) is conceptually part of the page's own lane, but the owned-surface globs in .workflow/rooms/*.html were drawn along directory boundaries (src/app/ vs tests/) rather than along the 'guards X' pairing the repo's own commenting convention already encodes — the first structural gap of this kind found in this room's drift history (prior three lessons here were facts-only, no mechanism named).

     The missing guard (the deliverable): A test that walks every `Guarded by tests/<path>` comment found inside a file under a room's owned-surface glob, and asserts that `<path>` is also covered by that same room's glob in .workflow/rooms/<room>.html — failing loudly the moment a page and its stated guard diverge in ownership.

     Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Saturated, not right-sized — 100% lit-to-walked ratio on both lanes means the walk floods to completion rather than decaying, so it adds no signal beyond the seed. Concrete adjustment: introduce ply-decay (per AR-2's own anti-rule) so a cell's match score must clear a floor to stay lit, restoring litIntent<walksIntent headroom and making the walk itself informative again. Experiments: Re-run this commit's ingest with tests/surfaces/*.test.mjs comment-headers included in the reef seed corpus for src/app pages, and measure whether prewalkOverlap rises above 2 and offPct drops below 15 · Sample the last 20 builder commits that touch a page+paired-guard pair together, and check whether prewalkOverlap<5 predicts offPct>15 reliably enough to flag the commit before the walk runs, as a cheap pre-filter · Add ply-decay to the walk and re-render this exact sha's panel to see whether litReality separates from walksReality once decay is active, confirming the walk-utilization hypothesis above

     Measured outcome (deterministic, not self-graded): room mean offPct 23→23 over 6 commits since · dictionary ablation saved 1603B

     The 22% off-lane reading here is a false-positive premium load on 🔨 builder's loss-ratio — the mass is a legitimate paired copy+guard edit, not genuine cross-domain drift — and adding tests/surfaces/*.test.mjs to the room's owned-surface glob removes that load without masking any future real encroachment.

     Open the signed attestation →

138. 2026-08-31 · 🔨 builder · offPct 22 · structural

     feat(iamfim): Determinism means chaos — Rice carries the fold, and the link goes live

     Why it left the lane: The diff is dominated by prose-craft rewrites to a single landing page (band copy, kicker link, claim wording) plus a matched test-regex fix — all legitimate 🔨 builder territory (src/app/, tests/surfaces/). But the reef seeds on ingested TERMS, and this commit's terms are saturated with theorem/insurance/underwriting vocabulary ('Rice', 'theorem', 'underwriter', 'price a number', 'countable: detected, placed, priced') that this repo's lattice anchors to the pricing/underwriting axis (B/C-side coordinates), not to the builder's UI-surface axis. The visit-shape route.ts lane-name string 'the kicker gag' and the founder-LinkedIn prose additionally drag toward the network/comms room's vocabulary (signature, handle, verified-vs-not). The walk saw dense semantic mass pointing at pricing-and-verification concepts sitting inside a builder-owned file, and spread wide (137 vs 131 lit cells) rather than converging tightly on 'UI copy edit' — which is exactly what offPct=22 measures: the room claim is right about the FILE, wrong about the CONCEPT DENSITY that file now carries.

     Intent–reality gap: The commit message frames this as a copy/craft pass on one page ('five operator corrections... on the same span'), and the room claim (🔨 builder) matches the files touched — but the actual prose landed carries load-bearing thesis vocabulary (Rice's theorem, underwriter pricing, verification epistemics) that the reef treats as its own axis, so the receipt sees reality mass outside the builder lane even though every line is in a builder-owned file.

     • src/app/iamfim-landing/page.tsx: add one sentence to the band's block comment naming the deliberate borrow of Rice/underwriting vocabulary from the canonical thesis, not new claims
     • commit message template (docs or CLAUDE.md COMMIT MESSAGE STRUCTURE section): add optional 'Carries-Thesis-Vocabulary:' trailer for builder-lane commits that quote canonical cross-axis claims
     • scripts/pmu/ (reef ingest): consider tagging known canonical-vocabulary spans (Rice, underwriter, determinism) so ingest can mark them 'cited canon' rather than 'novel claim mass' when computing lane placement
     • tests/surfaces/iamfim-determinism-is-reproducibility.test.mjs: the fix already lands construct-matching; add a comment noting the guard's scope is content-correctness, not lane-placement — it will not catch this drift class
     • docs/architecture/anti-rules-ledger.md or reef docs: note 'quoting the canon is not the same as introducing new lattice mass' as a named exception to how ingest scores concept density

     What we learned: Prior lessons for this room show offPct rising and falling with no visible mechanism (16→46, 15→16) with facts-only intervention; this is the first sensemaking-available pass and it reveals a NEW structural fact: a single-file, single-room commit can still trip lane threshold purely from VOCABULARY DENSITY of canonical cross-axis terms (Rice, underwriter, pricing) embedded in otherwise-correct UI prose — the reef does not distinguish 'quoting the thesis' from 'doing thesis work', so builder-lane content commits that reference the pricing/underwriting canon are structurally exposed to this drift class regardless of which files they touch.

     The missing guard (the deliverable): A canon-citation allowlist check: a script that scans a commit's page-content diff for known canonical thesis phrases (Rice, 'underwriter can price', 'the same commit, twice', 'two determinisms') and, when found inside a builder-owned UI file, asserts the commit message carries an explicit acknowledgment trailer (e.g. 'Carries-Thesis-Vocabulary:') — so citation-driven drift is distinguished from unbounded/undeclared cross-lane work at commit time rather than only after the receipt fires.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent(137) and litReality(131) are both healthy fill numbers relative to a 144-max ballistic run (walkTruncated=true caps it), so the walk is not under-using its budget — it's saturating near the ceiling on both sides. Adjustment: since walkTruncated=true, re-run this commit with a higher --max-depth or larger cap to see if the extra plies would have converged litReality toward litIntent (reducing offPct) or diverged further (confirming the drift is real, not a truncation artifact) — right now truncation makes it impossible to tell whether offPct=22 is genuine or a clipped-walk artifact. Experiments: Re-run sha 1fb1c3a8f's ballistic walk with walkTruncated disabled / max-depth doubled and compare offPct — measurable outcome: offPct either drops (truncation artifact) or holds/rises (real drift), settling the walk_utilization question above · Build a small canon-phrase allowlist (Rice, 'two determinisms', 'price a number', underwriter/underwriting) and re-ingest this commit's diff with those phrases masked/neutralized before reef seeding, then recompute offPct — measurable outcome: if offPct drops materially, canon-phrase density is confirmed as the dominant pull and the allowlist-and-mask approach is validated as a general reef preprocessing step · Compare prewalkOverlap across the room's last 10 out-of-lane commits vs its last 10 in-lane commits — measurable outcome: if out-of-lane commits systematically show prewalkOverlap <10 while in-lane commits show >20, that threshold becomes a cheap pre-walk predictor of drift risk, usable to flag commits before the full walk runs

     Measured outcome (deterministic, not self-graded): room mean offPct 27→10 over 6 commits since · dictionary ablation saved 1033B

     A single citation-driven drift in an otherwise-correct UI commit is a low-severity, high-frequency event for this lane (56 in 14 days already) — it erodes the room's loss-ratio not through bad code but through unexplained noise in the drift signal itself, and premium is restored by making canon-citation legible (the trailer + comment) so the reef can eventually discount cited vocabulary rather than scoring it as fresh cross-lane mass.

     Open the signed attestation →

139. 2026-08-31 · 🎤 voice · offPct 37 · mistagged

     feat(milestones): add "Countersign autonomous moves" — the tape has to outlive the mover

     Why it left the lane: The diff is entirely data/*.json (milestones-market-contact.json, milestones-next.json, milestones-progress.json, milestones.json) — a milestone-board asset edit. The commit message's prose is dense with infrastructure-signal terms: 'verifier that countersigns', 'exec predicates', 'record carries its commit', 'AXIOM 1's witness corollary', 'executions.ndjson', '.thetacog/elicit/' — all C.Operations/B-Tactics vocabulary (records, signals, countersigning, replay) with zero A.Strategy framing. The receipt's own walk (litReality=134 > litIntent=128, prewalkOverlap only 5) shows the message text pulled the reef into 31 out-of-role Operations×Tactics cells (Grid/Loop/Flow × Speed/Deal/Signal) that the actual changed files — four static JSON tables — never touch. The message is doing the semantic work of a design doc for a system (the exec/verifier predicates) that this commit does not implement; it only records the MILESTONE ENTRY describing that future work.

     Intent–reality gap: The commit message reads as an architecture/implementation narrative (verifier predicates, countersigning, replay, AXIOM 1 witness corollary) but the receipt shows a docsOnly=true, walkTruncated=true asset-only diff — four milestone JSON files. Nothing in the diff implements a verifier, a countersign predicate, or touches executions.ndjson; the commit only ADDS AN ENTRY describing that this work is needed.

     • Reword the commit-message body (already shipped as 4796c2cf12) is immutable — instead, log a correction note in the milestone entry's own JSON (e.g. an `implementation_status: "not started"` field in data/milestones.json for this tri-word) so the board itself, not just prose, marks the gap.
     • Add a `scope` or `entry_type: "board-only"` field to data/milestones-next.json / data/milestones.json schema so future milestone-add commits can self-declare docsOnly intent, letting the receipt's docsOnly flag be cross-checked against a declared field instead of inferred purely from diff shape.
     • When the actual verifier/countersign code lands in a future commit, its message should explicitly reference this milestone entry's sha (4796c2cf12) so the two-sided story (ticket → implementation) is linkable.
     • Add a lint in the milestone-add helper script (if one exists, e.g. scripts touching data/milestones*.json) that warns when a commit message's verb tense is imperative-present ('the record carries', 'a verifier signs') for a diff that only touches data/*.json — a cheap tense/diff-shape cross-check.

     What we learned: A milestone-board commit (pure data/*.json, docsOnly=true) is structurally the highest-risk shape for this drift class: the diff itself carries almost no semantic mass (four JSON files, no code identifiers), so the ENTIRE placement signal comes from the commit message's prose — and a milestone entry's rationale paragraph is written in the same register as a real implementation commit for the feature it's proposing. Unlike a code commit where the diff anchors the walk, a board-only commit lets the message alone drag the walk anywhere its vocabulary points, with nothing in the diff to counterbalance it. This is a distinct root cause from the prior three facts-only events in this room (which had no sensemaking recorded) — it's the first time the mechanism is legible: docsOnly commits are structurally undefended against message-driven drift.

     The missing guard (the deliverable): A guard that flags any commit where docsOnly=true AND diff touches ONLY data/milestones*.json AND the message body contains ≥N implementation-register terms (verifier, predicate, signs a verdict, the record carries, hence the exec) — these are future-tense/design claims embedded in present-tense/imperative prose for a commit that ships no code. The guard would warn (not block) at commit time: 'milestone-entry commit reads as implementation — confirm no code shipped, or move design prose to rationale/future-tense.'

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized in depth (2 plies, low fill 0.98%, well under the 70% saturation ceiling) but starved of intent-side anchors: litIntent=128 vs litReality=134 with prewalkOverlap only 5 means the two sides barely overlapped before the walk began, so the walk had almost no shared ground to converge on. Adjustment: for docsOnly=true commits, weight the intent-side seed more heavily toward the diff's own file paths (data/milestones*.json → C.Operations/board-vocabulary) rather than letting the full message text seed intent, since the message is the exact thing driving drift here. Experiments: Experiment: for commits where docsOnly=true and diff is 100% data/*.json, halve the message's contribution to intent-seed mass (or cap message tokens used) and re-run the receipt on this same sha (4796c2cf12) — measure whether offPct drops below 15 without losing legitimate signal, confirming the message-dominance hypothesis. · Experiment: add a synthetic intent anchor derived from the touched file's own JSON schema keys (e.g. 'milestone', 'tri-word', 'market', 'progress') for data/milestones*.json commits, then re-run the walk and measure whether litIntent rises closer to litReality's 134, tightening prewalkOverlap. · Experiment: run the last 20 docsOnly=true commits in this room's history through the same offPct measurement and check whether offPct correlates with message length/implementation-term density — if yes, that's a reusable pre-commit signal, not just this one event.

     Measured outcome (deterministic, not self-graded): room mean offPct 16→18 over 4 commits since · dictionary ablation saved 1101B

     An out-of-lane board-only commit at offPct 37 in a room already carrying 92 out-of-lane events in 14 days reads as sustained message-driven drift rather than a one-off, which erodes this lane's loss-ratio; premium is restored by the entry_type schema field plus the docsOnly-register guard turning this specific mechanism into a checkable, declining signal rather than an unexplained recurring one.

     Open the signed attestation →

140. 2026-08-31 · 📐 architect · offPct 24 · mistagged

     fix(funnel): the doors were at 87% depth — position was the defect, not content

     Why it left the lane: The diff is pure funnel-copy/UX work in src/app/hiring/page.tsx — 'If you sell', 'If you price risk', 'If you write code', 'fork it and send something back', 'explain the licence to someone who signs things' — plus a regex fix in tests/funnel/hiring-funnel-ratchet.test.mjs. None of that is docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, or CLAUDE.md — architect's entire owned surface. The commit's own Story trailer self-declares 'B.Tactics × C1.Operations.Grid', and Relevant-Rooms lists architect+operator+navigator together — the commit is honest about being cross-lane. The receipt still prices the whole thing against architect because Originating-Terminal (which room's terminal was physically open) is being read as lane truth, and 24% of the lit reality mass in a genuinely Tactics/Operations commit reads as off-lane for a Strategy-docs room.

     The semantic pull: Repeated funnel/sales vocabulary in the diff ('doors', 'homework', 'Step one/Step two', 'they are worth exactly the same to us', 'fork it and send something back') is Tactics×Operations vocabulary by the commit's own labeling — it pulls placement toward C1.Operations.Grid coordinates, away from anything docs/strategy/ would seed. No region narrative is cached for this sha, so this is read off the diff strings and the Story trailer's own coordinate claim, not off the encircled PNG.

     Intent–reality gap: The commit message claims a topology finding about the hiring funnel (accurate, and self-labeled Tactics×Operations) while the receipt attributes the whole event to 📐 architect's Strategy-docs lane via Originating-Terminal — the claim on the commit and the claim the detector is testing against are two different lanes.

     • Ship the unified `room-drift-<room>-check.sh` (or equivalent .mjs) that three rooms have each independently spec'd since ff09774e5 but nobody has shipped — this event is proof the gap is still open.
     • Update whichever script computes per-commit offPct to union Relevant-Rooms' owned globs against changed files instead of keying solely on Originating-Terminal.
     • Add tests/funnel/ and src/app/hiring/ to a Tactics/Operations room's (not architect's) owned-surface declaration in .workflow/rooms/*.html so a correctly-matching room exists for this class of commit.
     • Add a regression fixture using sha 4b9efa502d itself (real bytes, not invented) to the guard-seen-to-fail ledger once the detector fix ships, per the 'guard that has never been seen to fail' rule.
     • Extract the 3x-repeated door markup (`border-l-2 border-emerald-500/60 pl-4`, `text-emerald-300 font-semibold mb-2`, `text-slate-300 leading-relaxed`) into a shared `<EntryDoor>` component while touching this file again — unrelated to lane drift but visible boilerplate in this diff.

     What we learned: This is the third independent discovery of the same root cause (ff09774e5 named it; two other rooms spec'd it too) and the room's offPct trend across the three prior lessons is 14→28, then 17→26, and now this event — climbing, not falling. A sensemaking pass that produces another diagnosis without a shipped commit does not reduce recurrence for this class; the structural fix needs an owner and a merged artifact, not a fourth write-up of the same finding.

     The missing guard (the deliverable): tests/pmu-simulator/room-lane-uses-relevant-rooms.test.mjs — feed the detector a synthetic commit whose changed files match ONLY a listed Relevant-Room's owned globs and NOT Originating-Terminal's; assert the off-lane percentage is computed against the matching Relevant-Room, not against Originating-Terminal. Include sha 4b9efa502d's real file list (src/app/hiring/page.tsx, tests/funnel/hiring-funnel-ratchet.test.mjs) as the seen-to-fail fixture once the classifier is patched.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Right-sized in coverage (135 vs 127 lit cells, close counts) but walkTruncated=true is set, meaning it hit a depth/time budget before exhausting the graph. Given the already-low prewalkOverlap, a truncated walk may not have had room to discover the true overlap. Adjustment: rerun this sha with `--max-depth` increased and compare offPct — if it drops materially, truncation (not real drift) explains part of the 24%. Experiments: Re-seed the architect reef with Relevant-Rooms trailer text (architect, operator, navigator + the commit's own Tactics/Operations vocabulary) and recompute prewalkOverlap for this sha — expect it to rise above 4 if cross-room trailer awareness helps. · Rerun sha 4b9efa502d with an increased --max-depth to remove walkTruncated=true and diff the resulting offPct against 24 to isolate how much of the signal is a truncation artifact. · A/B the last 20 architect commits: compute offPct via Relevant-Rooms-union globs vs Originating-Terminal-only globs, and check whether the union method collapses known false positives (like this one) without hiding a known true positive.

     Measured outcome (deterministic, not self-graded): room mean offPct 24→7 over 3 commits since · dictionary ablation saved 880B

     A false-positive off-lane flag taxes architect's loss-ratio for cross-room work its own commit trailers already disclosed and CLAUDE.md already sanctions, inflating apparent drift with no real event behind it; premium is restored only when the detector prices lane against Relevant-Rooms' surface instead of Originating-Terminal alone — the fix this and two prior lessons have already priced but nobody has shipped.

     Open the signed attestation →

141. 2026-08-31 · 🎤 voice · offPct 34 · mistagged

     feat(milestones): add "Anchor third-party timestamps" — the one SETTLED claim that does not survive its own audience

     Why it left the lane: The diff's actual payload is data/milestones-*.json (2,269 of 1,562 diff lines) plus a one-comment addition to tests/milestones/milestones-derived.test.mjs — a VERBS allowlist gaining 'anchor' and 'serve'. The commit message is a legal/epistemic argument about third-party witness status (git hosting, force-push, tamper-evidence) and reads as governance/legal-audit prose, but the receipt only sees the reality corpus: identifier words and comments in milestone JSON + one test file. 'anchor' and 'serve' are generic English verbs with no repo-specific semantic gravity toward B3.Tactics.Signal (git-discipline) — they scatter across Tactics.Speed/Deal and Operations.Grid/Loop/Flow because those are the nearest lit neighbors to 'milestone board' content in the reef, not because the work is git-discipline work. The commit was placed in this lane by trailer/room convention (Originating-Terminal: 🎤 voice) rather than by what the diff's mass actually is: milestone-board data restructuring.

     The semantic pull: The 31 encircled out-of-role cells cluster on B3.Tactics.Signal(x11)/C1.Operations.Grid(x6)/C2.Operations.Loop(x4)/C3.Operations.Flow(x4) — exactly the Tactics×Operations quadrant, which is where 'milestone', 'settled', 'market', 'predicate', 'ledger' terms from data/milestones-*.json land when walked. The diff stat confirms it: 1121 lines in milestones-progress.json alone, 408 in milestones-market-contact.json, dwarfing the 6-line test comment that is the room's only claimed 'voice' territory (a VERBS allowlist edit, itself closer to C.Operations/B3.Signal than the JSON payload is).

     Intent–reality gap: The commit message frames this as a witness-corollary / legal-audit fix ('GitHub is hosting not witness', 'AXIOM 1 pointed at us') scoped to one requirement's status — narrow, room-appropriate prose for 🎤 voice. The receipt measures where the WORK landed: four milestone-data JSON files carrying 2,269 changed lines of board content, none of which the commit message names or bounds.

     • Confirm data/milestones-progress.json etc. are GENERATED from milestones.json (grep for the generator script per NEVER WRITE STATE INTO A GENERATED FILE) — if so, note the generator name in the commit message per that HARD RULE.
     • Add a `derived-from: data/milestones.json` comment or manifest key atop each of the four large derived files so future receipts and humans can see the causal chain without diffing 1,121 lines.
     • tests/milestones/milestones-derived.test.mjs: add an assertion that the four derived JSON files are internally consistent with milestones.json's new 'anchor'/'serve' entries (a regeneration check, not just a VERBS-set check).
     • Split future milestone-board edits into two commits: (1) the authored requirement change to milestones.json + VERBS test, (2) the mechanical re-render of the derived files — restores one-iteration-one-commit and gives each its own honest lane.

     What we learned: A commit whose message is entirely legal/epistemic reasoning about ONE requirement can still be, by diff mass, a full-board regeneration — the room-tagging convention (Originating-Terminal) tracks WHO wrote the prose, not WHERE the bytes landed, and derived-file regeneration (milestones-progress/market-contact/next/causal.json) is the single largest untagged source of offPct in this room's last 96 events.

     The missing guard (the deliverable): tests/milestones/milestones-derived.test.mjs (or a new pre-commit check): when a commit touches milestones.json AND any of milestones-{progress,market-contact,next,causal}.json in the same commit, require the commit message to contain a 'Touches:' or 'Derived:' line naming the derived files explicitly — cheap, mechanical, and it would have caught this commit's unbounded message before it shipped.

     Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. litReality(133) > litIntent(107) with walkTruncated=true at only 2 plies — the walk is UNDER-DEPTHED for a 1,562-line/6-file diff: it's truncating before the recursive row→column→row chain can concentrate reality back toward the intent lane. Concrete adjustment: raise --max-depth for commits whose diff spans >3 files / >500 lines (this one qualifies on both) so the walk gets the plies needed to resolve the derived-file mass back to its milestones.json source rather than stopping mid-spread. Experiments: Re-run this exact sha's receipt with --max-depth 4 (vs the truncated 2) and compare offPct — if it drops, walkTruncated is the dominant cause and depth-budgeting by diff size becomes a standing rule. · Ablate: exclude the four milestones-*.json files from the reality corpus for this sha only (treat them as derived/non-authored) and recompute offPct — tests whether a 'derived-file' corpus flag would have kept this in-lane. · Compare offPct for this sha against a synthetic commit that touches ONLY milestones.json + the test file (no derived JSON) — establishes the baseline offPct for 'the actually-authored change' vs the full commit.

     Measured outcome (deterministic, not self-graded): room mean offPct 18→15 over 4 commits since · dictionary ablation saved 740B

     One mistagged multi-file board regeneration nudges 🎤 voice's loss ratio up incrementally (96 prior out-of-lane events already show this is a chronic pattern, not a one-off), but it's cheap to restore: the Touches: trailer guard plus a corrected owned-surface glob converts 'unbounded board rewrite' into 'declared derived re-render,' which is exactly the kind of bounded, re-runnable claim underwriting can price.

     Open the signed attestation →

142. 2026-08-31 · 🎤 voice · offPct 18 · mistagged

     fix(milestones): revive the dark causal generator — export the ref() it links against

     Why it left the lane: The diff's actual vocabulary — `derive.mjs`, `export function ref(item)`, `causal generator`, `data/milestones-causal.json`, `the Node linker itself computes`, `tests/milestones/no-dark-scripts.test.mjs` — is Operations/tooling language with zero lexical overlap to 🎤 voice's owned surface (books/tesseract/chapters, src/content/blog/*.mdx, decentralizedtalk/slideData.ts, src/app/6needs/, scripts/voice-*-check.sh). No region narrative is cached for this sha, but the room-lane mismatch is visible without it: the commit's own trailers already disagree — `Originating-Terminal: 🎤 voice` while `Relevant-Rooms: 🎩 operator, 🔨 builder, 📐 architect` names three other rooms and omits voice entirely. The reef placed the reality mass where the text actually compresses closest — Operations/builder coordinates — while the receipt scored it against voice's lane because that's the terminal the commit was authored from.

     The semantic pull: Terms pulling toward Operations/builder: `derive.mjs`, `ref()`, `SyntaxError`, `set difference`, `the Node linker`, `dark script`, `guard`. None of these, nor the two touched files (scripts/milestones/derive.mjs, tests/milestones/no-dark-scripts.test.mjs), sit anywhere near voice's book/blog vocabulary — the pull is total, not partial, which is consistent with a genuine cross-room commit rather than a borderline one.

     Intent–reality gap: The commit message and the diff are internally coherent — it does exactly what it says, fixing a dead generator and adding a linker guard. The gap is not between claim and code; it's between the room the receipt scored this against (🎤 voice, from Originating-Terminal) and the room the work actually belongs to, which the commit's own Relevant-Rooms trailer already names correctly (operator/builder/architect).

     • Add the Originating-Terminal-vs-Relevant-Rooms divergence rule to CLAUDE.md's COMMIT MESSAGE STRUCTURE section (the spec_rewrite above)
     • Add scripts/milestones/**, tests/milestones/**, data/milestones*.json to an explicit owned-surface glob in the operator (or architect) room's .workflow/rooms/*.html — currently likely undeclared, which is why the lane-scorer had nothing to fall back on but the terminal
     • Confirm whether the receipt/lane-scoring pipeline consults Relevant-Rooms at all today, or only Originating-Terminal — this event suggests the latter
     • Check the other 90 out-of-lane 🎤 voice events from the last 14 days for the same Originating-Terminal∉Relevant-Rooms signature — if it recurs, the terminal-based scoring is the room's dominant failure mode, not content drift

     What we learned: The lane-scoring pipeline appears to key a commit's room purely off Originating-Terminal (which physical terminal was open), not off the commit's own Relevant-Rooms self-report or the touched-file owned-surface globs — so a commit can openly name three other rooms as more relevant in its own trailer and still get scored as an out-of-lane event for the terminal room, with the self-contradiction never consulted. This is different from the prior three voice-room lessons (all facts-only, no diagnosed mechanism) — this is the first time the mechanism is nameable: it's a scoring-input choice, not a content-drift problem.

     The missing guard (the deliverable): A check — advisory at minimum — that flags when a commit's Originating-Terminal does not appear in its own Relevant-Rooms list (e.g. tests/ops/originating-terminal-in-relevant-rooms.test.mjs, or a warn-only hook alongside the existing trailer conventions), so a cross-room commit is visible before it silently scores as a same-room drift event. Stronger version: feed that divergence into the receipt generator itself so lane-scoring for a flagged commit uses the union of Relevant-Rooms' owned-surface instead of Originating-Terminal's.

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. Saturated, not shallow: every walk that ran lit (118/118 intent, 111/111 reality) — the walk isn't dying early or thinning out, it's fully exploring what it reached. The fix isn't depth, it's the seed: re-seed the walk from Relevant-Rooms vocabulary (operator/builder/architect terms) in addition to Originating-Terminal's, since this commit's own trailer already names the better seed. Experiments: Re-render this sha's receipt seeding the reef from Relevant-Rooms (operator+builder+architect) instead of Originating-Terminal (voice) alone; measure whether offPct drops under 15 · Add scripts/milestones/**, tests/milestones/**, data/milestones*.json to an explicit owned-surface glob under the operator or architect room, then re-render and check whether offPct moves — tests whether owned-surface globs are consulted by the pipeline at all or are dead declarations · Across the room's other 90 out-of-lane events in the last 14 days, compute the correlation between (Originating-Terminal ∉ Relevant-Rooms) and offPct>threshold; a strong correlation validates trailer-mismatch as a cheap leading indicator the scorer could check before rendering

     Measured outcome (deterministic, not self-graded): room mean offPct 16→18 over 4 commits since · dictionary ablation saved 797B

     A mistagged-room event like this one is a false positive against 🎤 voice's loss ratio — it inflates the room's out-of-lane count without any actual content drift — so premium correction here means fixing the attribution input (spec_rewrite + guard), not tightening voice's actual content discipline, which was never at issue.

     Open the signed attestation →

143. 2026-08-31 · 📐 architect · offPct 43 · mistagged

     fix(funnel): correct three probes against ground truth — verdict PASSES at floor 95

     Why it left the lane: The touched file, scripts/funnel/aggregate.mjs, sits outside every glob in architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) — there is no scripts/ entry at all. The diff's own vocabulary reinforces the pull away from Strategy/docs: PROBE regex bodies are saturated with sales/deal/money terms — 'commission|percentage|cut of|upside|earnings|income|paid|\$\s?20|agent-year|no salary|invoice|lucrative', 'selling (task|role|path)|sales homework|run(ning)? the demo.{0,80}(licen[cs]e|employer|buyer)' — plus persona/action terms like 'hacker', 'fork', 'open a pull', 'wouldAct'. None of that is strategy-memo prose; it's funnel-instrumentation engineering. Tellingly, the commit's own Story trailer self-labels the work 'C1.Operations.Grid' — the room diagnosed its own coordinate correctly in prose while the Originating-Terminal trailer still says architect, which is exactly the mismatch the receipt measured as offPct 43.

     The semantic pull: Money/sales/persona regex vocabulary (commission, licen[cs]e, employer, buyer, hacker, fork, pull) plus a file path with zero overlap to any architect-owned glob drags the walk's lit reality mass toward an Operations/Tactics-flavored region, away from the docs/strategy lane the Originating-Terminal trailer claims. Region narrative wasn't cached for this sha, so this is inferred from the diff vocabulary and the Story trailer's self-declared coordinate rather than from named encircled coords.

     Intent–reality gap: The commit message and Originating-Terminal trailer file this as architect-room work, but the receipt measured 43% of the lit reality mass landing outside architect's declared lane — consistent with the actual file (scripts/funnel/aggregate.mjs) never having belonged to architect's owned surface in the first place, and the work being self-described in the Story field as C1.Operations.Grid, a different axis entirely.

     • Ship the lane-classifier union fix (Originating-Terminal ∪ Relevant-Rooms owned surfaces) that three prior interventions (architect/navigator/laboratory) each spec'd as room-drift-<room>-check.sh but none shipped — this is the actual deliverable, not another diagnosis.
     • Add scripts/funnel/** to the owned-surface glob of whichever Relevant-Room (🧭 navigator or 🎩 operator) actually claims funnel-pipeline code — right now it belongs to none.
     • Add one line to CLAUDE.md's COMMIT MESSAGE STRUCTURE section stating explicitly that offPct is computed against the union of Originating-Terminal + Relevant-Rooms surfaces, not Originating-Terminal alone.
     • Add a regression test asserting a commit whose Originating-Terminal differs from its Relevant-Rooms, but whose touched files match a Relevant-Room's owned surface, does NOT trip the offPct alarm.
     • File a bifurcate task naming the room that should own scripts/funnel/ so it stops landing in an unclaimed no-man's-land on every funnel-probe fix.

     What we learned: This is the third independent diagnosis of the identical Originating-Terminal-vs-Relevant-Rooms lane gap (after ffaf8d2ac/ff5110f85's unnamed facts-only passes and ff09774e5's named one), and the fix it named has STILL not shipped — room mean offPct only moved 20→14 before drifting back up (this commit sits at 43). The durable fact this reveals is not about the lane logic itself, which is now well understood, but about the intervention pipeline: it is capable of re-diagnosing a known, already-specified defect indefinitely without ever forcing the shipped repair, because each pass is scored on producing a good narrative rather than on closing a previously-opened ticket. The missing piece is a mechanism that checks whether a NEW event matches a PRIOR unshipped lesson and escalates to 'ship it now' rather than accepting a fourth restatement.

     The missing guard (the deliverable): tests/pmu-simulator/lane-classifier-honors-relevant-rooms.test.mjs — asserts the offPct computation for a commit is measured against the union of Originating-Terminal's and every Relevant-Room's owned-surface globs, and fails today because no such union exists; paired with a meta-guard that flags when an intervention's `lesson` text substantially repeats an unresolved lesson from the same room's history (via gzip-NCD against prior lesson strings) without a corresponding shipped file in the same or next commit.

     Detector probe: signal attribution — gzip reef 35% · ballistic walk 65%. walkTruncated=true on a commit this size (11-line diff) suggests the walk is being cut short rather than saturating naturally — a truncated walk on the reality side (129 lit, still climbing) versus a non-truncated-looking intent side (116) can manufacture asymmetry that reads as drift when it may just be an incomplete measurement. Concrete adjustment: bump --max-depth for commits under this diff-size threshold and re-run to see whether offPct 43 collapses once both sides finish walking. Experiments: Re-render this sha's receipt with --max-depth raised until walkTruncated=false on both sides; compare the resulting offPct against 43 to quantify how much of this event is a truncation artifact. · Recompute offPct for the same commit using union(Originating-Terminal, Relevant-Rooms) owned-surface as the lane definition and re-score all 186 architect out-of-lane events from the last 14 days to measure how many collapse under the 15 threshold. · Seed the reef dictionary with this commit's persona/probe vocabulary (ownDoor, wouldAct, inverted, refusedPhrase) and re-measure prewalkOverlap before/after to test whether funnel-probe commits are structurally sparse-seeded regardless of lane correctness.

     Measured outcome (deterministic, not self-graded): room mean offPct 24→25 over 6 commits since · dictionary ablation saved 589B

     186 out-of-lane events in 14 days inflates this lane's apparent loss ratio with a known, three-times-diagnosed-but-unshipped false-positive class (Originating-Terminal-only lane scoring), so premium stays mispriced until the union-lane fix and the scripts/funnel/ ownership assignment actually ship rather than being re-diagnosed a fourth time.

     Open the signed attestation →

144. 2026-08-31 · 📐 architect · offPct 24

     fix(funnel): withdraw the score trend — n=1 cannot see an edit, and I edited off noise twice

     Why it left the lane: (sensemaking unavailable — facts-only intervention)

     Measured outcome (deterministic, not self-graded): room mean offPct 23→12 over 6 commits since

     Open the signed attestation →

145. 2026-08-31 · 📐 architect · offPct 16 · structural

     feat(hiring): four crew hooks, one per visitor, and the click becomes countable

     Why it left the lane: The diff's real mass is src/lib/crewHooks.mjs (hookFor/hrefFor/pickCrewKey/CREW_HOOKS), src/components/CrewBar.tsx (useEffect/localStorage/Storage wiring) and tests/hiring/crew-bar-rotates.test.mjs — pure implementation + test vocabulary, none of which falls inside 📐architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, docs/01-business/strategy/, CLAUDE.md). The commit's own trailers already say this: Relevant-Rooms lists 🎩operator and 🔨builder alongside 📐architect, but the receipt's intent lane is keyed off Originating-Terminal alone (📐architect), so the walk's reality mass over code+test files reads as 16% off that single-room lane even though the commit itself named the correct rooms for that content.

     The semantic pull: Identifiers like `hookFor`, `pickCrewKey`, `Storage`, and the file paths `src/lib/crewHooks.mjs` / `src/components/CrewBar.tsx` / `tests/hiring/crew-bar-rotates.test.mjs` are exactly the code/build-surface vocabulary that gzip-NCD would seed toward builder/operator coordinates, while the intent lane fixed on architect carries strategy/doc vocabulary (docs/strategy, docs/03-product). No region narrative is cached for this sha, so the exact [row,col] anchors aren't available — the mismatch is inferable from the owned-surface comparison alone, which is sufficient to place the class.

     Intent–reality gap: The commit claims 📐architect as its Originating-Terminal while the substantive work — a new lib module, a component rewrite, and a 224-line Jest guard — is builder/operator territory; CLAUDE.md explicitly sanctions this Originating-Terminal vs Relevant-Rooms divergence as normal, but the receipt still measures drift against Originating-Terminal alone.

     • Ship (not just re-spec) `scripts/pmu/room-lane-check.mjs` — the union-of-Relevant-Rooms classifier three rooms already independently designed per lesson ff09774e5.
     • Add `src/lib/crewHooks.mjs`, `src/components/CrewBar.tsx`, and `tests/hiring/**` to 🔨builder's owned-surface glob in `.workflow/rooms/builder.html` so future receipts recognize this content natively.
     • Wire the new classifier into the same code path that emits the per-commit receipt row (sha/offPct/threshold) so the fix lands in the measured pipeline, not just as a doc.
     • Backfill-check the 187 out-of-lane events already logged for this room in the last 14 days against the union-lane rule to see how many were mistagged rather than genuinely off-lane.
     • Cross-link this fix in the other two rooms' (navigator, laboratory) copies of the same spec so the three independent designs converge on one shipped script instead of three unshipped ones.

     What we learned: This is the third independent occurrence of the exact ff09774e5 finding (architect, navigator, laboratory each hit it), and the room's own offPct trend (20→14, still nonzero) proves the diagnosis alone doesn't fix the pipeline — a correctly-tagged, CLAUDE.md-compliant commit (Originating-Terminal + Relevant-Rooms both present and accurate) still reads as 16% out-of-lane because nobody has shipped the union-of-Relevant-Rooms classifier despite three rooms independently designing it. The lesson is procedural, not architectural: a spec repeatedly re-derived without a single owner to ship it just keeps re-costing offPct against every room that touches multi-room commits.

     The missing guard (the deliverable): tests/pmu-simulator/room-lane-honors-relevant-rooms.test.mjs — feed the classifier a commit whose Originating-Terminal is room A but whose changed files match room B's owned-surface glob, with Relevant-Rooms correctly naming room B; assert offPct is computed against the UNION (so it does not flag drift), then assert a commit with NO Relevant-Rooms trailer covering the touched files still flags correctly (so the guard can't be satisfied by disabling the check).

     Detector probe: signal attribution — gzip reef 65% · ballistic walk 35%. litIntent/walksIntent = 142/142 and litReality/walksReality = 143/143 — both sides are fully saturated (every walked cell lights up), which combined with walkTruncated=true suggests the walk hit its depth/time budget before it could narrow past full coverage. Adjustment: raise --max-depth on this class of commit (multi-file, multi-room) so the walk can decay past ply 1 saturation and actually discriminate in-lane from out-of-lane cells instead of lighting nearly everything it touches. Experiments: Re-run this commit's receipt with the intent lane set to the union of architect+builder+operator owned-surface globs and measure whether offPct drops below 15 — validates the spec_rewrite before shipping the classifier. · Re-run with --max-depth raised by one ply and compare litIntent/litReality saturation ratios — if they drop below 1.0, the walk was under-budgeted; if they stay at 1.0, the saturation is structural to this reef, not a depth artifact. · Seed the reef dictionary with the crew-hook module's own vocabulary (hookFor/hrefFor/pickCrewKey/CREW_HOOKS) and re-measure prewalkOverlap — if it rises meaningfully, low overlap here was a vocabulary-coverage gap rather than a genuine cross-lane signal.

     Measured outcome (deterministic, not self-graded): room mean offPct 24→29 over 4 commits since · dictionary ablation saved 645B

     A structurally mistagged 16%-off event on a correctly multi-room-trailered commit inflates this lane's apparent loss ratio with false signal rather than real drift, and restoring it to true requires shipping (not re-spec'ing a fourth time) the union-of-Relevant-Rooms classifier already designed three times over.

     Open the signed attestation →

146. 2026-08-31 · 📐 architect · offPct 17 · mistagged

     fix(funnel): the harness was grading a reader who does not exist — skim window, and two real edits

     Why it left the lane: The diff lands entirely in scripts/funnel/qwen-read.mjs, docs/funnel/hiring-funnel-ratchet-v1.md and src/app/hiring/page.tsx. The room-ownership registry (.workflow/rooms/*.html) already has an opinion about that ground: 'funnel' is a listed OWNED keyword for both 🧪 cursor-laboratory.html and 🎭 alacritty-performer.html, and 'hiring' appears as a 'Next →' item for 🎭 performer and 🔨 iterm2-builder.html — never as owned or next for 📐 architect, whose own Owned list is benchmark/glossary/pmu. Confirming this isn't theoretical: cursor-laboratory.html shows THREE commits from 2026-08-31 doing this exact class of work under its own name — fb14220920 'run the qwen ratchet for real', a8431083bd 're-sequence /hiring on measured ghost-reads', 20a4c01561 'aggregate N reader samples into rates'. This commit (7f46d2ab7) is the same functional work, same day, tagged Originating-Terminal 📐architect instead.

     The semantic pull: The pulling terms are the file paths themselves, not prose: 'scripts/funnel/qwen-read.mjs', 'docs/funnel/hiring-funnel-ratchet-v1.md', 'src/app/hiring/page.tsx' — each matches laboratory's or performer's declared keyword ('funnel', 'hiring') rather than anything on architect's owned list (benchmark/glossary/pmu). No lattice region narrative was cached for this sha, so the anchor here is the room-registry keyword match, which is the mechanism this detector actually keys off — the same substance the missing cached region would have shown positionally.

     Intent–reality gap: The commit's trailers claim 📐architect (with 🎩operator/🧭navigator as Relevant-Rooms) doing Story-labeled 'C1.Operations.Grid' sensemaking work, but the receipt measured code landing squarely on ground the registry had already assigned elsewhere before this commit ran — laboratory owns 'funnel', performer/builder have 'hiring' staged as next. The claim isn't wrong about the WORK; it's wrong about whose work it is.

     • Execute the already-queued self-prompt sp-372 (due 2026-08-31, alacritty-performer.html): 'test update to ensure lane classification accuracy based on content globs' — this event is exactly its target class, don't let it lapse past its due date.
     • Execute the adjacent queued self-prompt sp-367: 'Confirm network's owned-surface glob covers necessary paths, update keywords, patch receipt generator, re-render' — same root mechanism, different room.
     • Resolve the duplicate 'funnel' ownership claim: both cursor-laboratory.html and alacritty-performer.html list 'funnel' under Owned. Pick one canonical owner or mark it explicitly shared, or the lane classifier has two right answers and no way to pick.
     • Add 'hiring' as an actual Owned keyword somewhere (it currently only exists as a 'Next →' aspiration on performer and builder, never a settled owner) so a hiring-page commit has a real home to be scored against.
     • Backfill: check whether fb14220920 / a8431083bd / 20a4c01561 (laboratory's same-day funnel commits) and 7f46d2ab7 (this one) represent two terminals doing the same work in parallel without coordination — a mesh/dispatch question, not just a labeling one.

     What we learned: The room-ownership registry is not a gap here — it already correctly assigns 'funnel' to laboratory/performer and stages 'hiring' as their next work, unlike the prior ff09774e5 case where the detector flagged behavior CLAUDE.md itself sanctioned. This is the opposite failure: a correct map sits in .workflow/rooms/*.html and the commit-authoring step never consults it when writing its own trailers, so a mistag happens even though the answer was already written down. Two rooms (architect and laboratory) worked the identical 'funnel'/'hiring' surface on the same day from different terminals with no visible coordination, which the registry alone cannot prevent — it can only make the mislabel visible.

     The missing guard (the deliverable): A warn-only pre/post-commit check (implements the already-queued sp-372) that reads the staged file paths, matches them against every .workflow/rooms/*.html 'Owned'/'Next →' keyword list, and prints a warning when Originating-Terminal names a room whose Owned list does NOT contain the best keyword match — mirroring scripts/ops/rules-for.mjs's pattern (surface the sibling rule the author didn't search for) but applied to room trailers instead of CLAUDE.md sections.

     Detector probe: signal attribution — gzip reef 20% · ballistic walk 80%. Right-sized but likely truncated short on the reality side: walkTruncated=true and litReality (127) sits ~6% below litIntent (135) despite both converting 100% of their walks to lit cells — a truncation cutoff before reality fully spreads would produce exactly this asymmetry. Adjustment: re-run this sha's reality-side walk with --max-depth increased by one ply and check whether litReality closes the 8-cell gap to litIntent. Experiments: Seed the reef dictionary with the .workflow/rooms/*.html Owned/Next keyword lists (funnel, hiring, benchmark, glossary, pmu, etc.) before running this sha's walk, and measure whether prewalkOverlap rises above 5 and offPct drops — tests whether the existing registry, once fed into the reef, would have caught this mistag pre-walk. · Re-run this sha with --max-depth +1 on the truncated (walkTruncated=true) reality walk and diff offPct against the current 17% baseline, to separate how much of the drift is a depth-budget artifact versus a real placement signal. · Run the sp-372/sp-367 guard (once shipped) retroactively over this room's 185 out-of-lane events from the last 14 days and report what fraction are funnel/hiring-class trailer mistags versus genuinely novel drift classes — sizes the blast radius before calling the fix complete.

     Measured outcome (deterministic, not self-graded): room mean offPct 24→25 over 6 commits since · dictionary ablation saved 798B

     This is a labeling loss, not a quality loss — the funnel-harness fix and hiring-page copy are sound — so it dents this room's rolling offPct trend (already 185 events/14 days) without reflecting bad work, and premium is restored the moment the Originating-Terminal/Relevant-Rooms trailer is corrected against the registry that already had the right answer, with no code change required.

     Open the signed attestation →

147. 2026-08-31 · 🎤 voice · offPct 18 · mistagged

     feat(milestones): add "Recompute sealed receipts" — the rung the whole board waits on

     Why it left the lane: The diff is 6 lines of test-file edit (adding 'recompute','countersign','form' to a VERBS allowlist), but the commit message is a ~250-word essay about milestone causal-graph state, market predicates, /api/receipt/verify's four claims, and RECOMPUTE_NOT_IMPLEMENTED — vocabulary from Operations.Grid/Loop/Flow (milestone board, market checks, execution bounds) and Tactics.Deal/Speed (exchange, prior handoffs, prerequisites). The reef seeds on the MESSAGE, not the 6-line diff, so the walk lands where the prose talks about — 31 encircled cells fanning across B1/B2/C1/C2/C3 — while the actual reality (a test allowlist edit) barely lights B3.Tactics.Signal at all. The diff-stat itself confirms the mismatch: 1435/1324 lines touched across 5 milestone data JSON files, and only 8 lines in the one file shown in the bounded diff — the receipt is measuring against a commit whose real weight is in files never surfaced to the walk.

     The semantic pull: 'causal graph' / 'prior handoff' / 'seven of eight tracks' / 'market predicates' / 'RECOMPUTE_NOT_IMPLEMENTED' / 'sealed artifacts' / 'recompute a receipt' are all milestone-board and market-verification vocabulary — they anchor to C1.Operations.Grid (the board), C2.Operations.Loop (the handoff sequencing), C3.Operations.Flow (the check pipeline), and B2.Tactics.Deal (the market/exchange framing) — not to B3.Tactics.Signal, which is where a VERBS-allowlist-in-a-test-file commit should land.

     Intent–reality gap: The commit message claims to be giving 'the rung the whole board waits on' — a milestone/causal-graph capability gap closed at the API level (recompute verdict) — but the actual diff shown is 8 lines in a test file's VERBS Set, plus (per the diff-stat) ~2,759 lines of data/milestones-*.json churn that never appears in the bounded diff at all. The receipt measured the prose's claim against a code surface the message barely describes.

     • Split future milestone-regen + VERBS-allowlist edits into two commits per ITERATION DISCIPLINE — the causal-graph JSON regen is its own iteration from the test-allowlist edit
     • Confirm data/milestones-causal.json, milestones-market-contact.json, milestones-progress.json, milestones-next.json, milestones.json are the actual diff surface the receipt pipeline ingests for this sha — if the bounded diff tool truncated them, the pipeline's ingest-scope needs to include milestone data files, not just 'semantic files first'
     • Add a milestones/ path check to the room-surface glob for 🎤 voice — voice does not own data/milestones-*.json (per OWNED FILE SURFACE above: books/blog/decentralizedtalk only) so this commit's real weight belongs to whichever room DOES own milestones data (builder/operator per Relevant-Rooms trailer) — the QC line already routes to 🔨 builder, which is correct, but the offPct receipt fires on the ORIGINATING room
     • Cross-check tests/milestones/milestones-derived.test.mjs's own docstring convention (the comment already narrates each verb's justification inline) — that comment IS the spec; commit message should point at it rather than re-deriving the same rationale in prose

     What we learned: A commit's message length/scope and its bounded-diff size can diverge by two orders of magnitude (250-word causal-graph essay vs 8-line test edit) when a single commit bundles a large generated-data regen with a small hand-written code change — the receipt pipeline's 'semantic files first' diff-bounding surfaces the small file and scores drift against a narrative written for the large one. This is a NEW class relative to prior voice-room lessons (which were facts-only, no diagnosed mechanism): the failure isn't the verb choice or the room's habitual off-lane pull, it's that ONE commit's message can legitimately describe TWO different-sized changes and the pipeline only ever sees one of them.

     The missing guard (the deliverable): A pre-commit check that flags when a commit's message names data files (milestones-*.json, causal graph, market predicates) that are NOT actually staged, OR conversely when staged data-file line-count vastly exceeds the code-file line-count the message narrates — e.g. `tests/comms/message-diff-proportionality.test.mjs`-style: if >70% of changed lines are in files never mentioned in the message body, warn that the message's claimed scope doesn't match its diff.

     Detector probe: signal attribution — gzip reef 75% · ballistic walk 25%. Right-sized on ply count (2 plies, 4.2k w/s, fill 0.98% — nowhere near saturated) but walking from a BAD seed: litIntent/litReality are nearly equal (130/132) while offPct is 18, meaning the walk faithfully reproduced a seed-level mismatch rather than a walk-depth problem. Adjustment: no depth change needed — the fix is upstream at seed/ingest (bound the message text fed to gzip-NCD to the actual diff's file list before seeding). Experiments: Run this same sha through the pipeline twice: once seeding gzip-NCD from the full commit message, once from a message truncated to only sentences containing tokens that appear in the diff's changed files (VERBS, allowlist, derived-test) — measure whether offPct drops toward 0 in the second run, confirming the message-vs-diff-scope hypothesis · Re-ingest this commit with the milestones-*.json files actually included in the walked diff (currently only the 8-line test file was shown as 'bounded, semantic files first') and check whether litReality/walksReality shift toward C1/C2/C3 — if the JSON files were excluded from ingest by the semantic-file filter, that filter itself may be the root of the offPct spike, not the message · Compute breach-rate (offPct>15) specifically for commits where diff-line-count-in-largest-file / diff-line-count-in-message-topic-file > 50x, across the last 120 commits, to see if this ratio predicts offPct independent of room

     Measured outcome (deterministic, not self-graded): room mean offPct 16→18 over 4 commits since · dictionary ablation saved 878B

     An 18%-off-lane receipt on a commit whose message essay outweighs its shown diff by ~300x is a message-hygiene loss, not a code-correctness loss — the lane's premium is restored by the two-commit split (checklist item 1) plus the proportionality guard, both of which are cheap, mechanical, and don't touch the underlying milestone/test logic that is itself correct.

     Open the signed attestation →

148. 2026-08-31 · 📐 architect · offPct 19 · mistagged

     feat(blog): the determinism trap — the phrase survives, the word you meant does not

     Why it left the lane: Every changed path in this commit — docs/05-content/blog/cook-rounds/2026-08-31-the-determinism-trap.{claims,predictions}.md and src/content/blog/2026-08-31-the-determinism-trap.mdx — is blog-authoring surface. The diff's own vocabulary ("Fork A"/"Fork B", "six needs in canonical order", "authority last, evidence-last digestif", the ten-course maître d' headers, the T1/T2 predicted-monologue sidecar) is voice/comms-pipeline jargon, not architect jargon. 📐 architect's actual owned reef terms (read from .workflow/rooms/vscode-architect.html) are benchmark / glossary / pmu / book / patent / docket / deck / dentry / skill — none of which the diff's content touches. The walk lit 127 reality cells against 126 intent cells at near-1:1 saturation (litIntent≈litReality, walksIntent≈walksReality), but only 45 of those cells pre-existed in the reef seed overlap before the walk ran — so the mass is landing on legitimate, well-formed content coordinates that simply sit outside the architect lane the Originating-Terminal trailer declared.

     The semantic pull: The prewalkOverlap=45 vs ~126-127 lit cells (≈35% seeded) shows the reef recognized this as strongly-formed content (blog/voice vocabulary: 'maître d'', 'course', 'digestif', 'ghost-read', 'monologue') that has almost no shared seed with architect's benchmark/pmu/patent/deck vocabulary — so nearly two-thirds of the lit mass had to be freshly discovered by the walk rather than confirmed against the declared lane, and it landed squarely in voice-owned territory.

     Intent–reality gap: The commit message claims architect-room work (Originating-Terminal: 📐 architect) and CLAUDE.md explicitly sanctions Originating-Terminal/Relevant-Rooms divergence as normal — but the receipt classifies lane by Originating-Terminal alone, so a commit correctly labeled 'physically committed from the architect terminal, content belongs to voice+operator' (which the commit's own Relevant-Rooms trailer already says) still measures as 19% out-of-lane against architect's owned surface.

     • Ship `scripts/ops/room-lane-classify.mjs` as the single canonical implementation (kills the 3 duplicate unshipped specs)
     • Add `src/content/blog/**` and `docs/05-content/blog/cook-rounds/**` to terminal-voice.html's Owned list explicitly
     • Add an exclusion note to vscode-architect.html: blog/content authoring is never architect's lane regardless of Originating-Terminal
     • Add one sentence to CLAUDE.md's COMMIT MESSAGE STRUCTURE section requiring Relevant-Rooms-union lane classification (see spec_rewrite)
     • Add a firing-test pair per the GUARD-SEEN-TO-FAIL rule: 84c4e13b8 (must go red under Originating-Terminal-only, green under union) as the historical defect fixture
     • Re-run the 189 architect out-of-lane events from the last 14 days against the union classifier and report how many resolve to in-lane

     What we learned: The root cause here was already correctly diagnosed once (ff09774e5, three rooms independently spec'd the identical fix) and STILL recurred — which reveals a new structural fact beyond the original diagnosis: a correct root-cause finding with no forcing function to ship it gets silently rediscovered rather than acted on, and nothing in the pipeline currently detects when ≥2 rooms independently propose the same unshipped fix as a signal that it should be prioritized and unified.

     The missing guard (the deliverable): A duplicate-spec detector: scan open specs/decisions across rooms (gzip-NCD near-duplicate match over spec text) and auto-flag when ≥2 rooms independently propose the same fix without either shipping it — this is the guard that would have caught `room-drift-<room>-check.sh` being spec'd three times and shipped zero.

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Too shallow relative to content size: walkTruncated=true on a 230-line, docsOnly=true blog post with 305 inserted lines means the walk hit its depth/time budget before naturally converging, right as litIntent/litReality were already near-saturating a 144-cell lattice (126-127/144 ≈ 87-88%). Concrete adjustment: raise --max-depth (or the time budget) specifically for docsOnly commits above some line-count threshold, then re-run 84c4e13b8 and check whether offPct changes purely from walkTruncated flipping to false. Experiments: Re-run the ballistic walk on 84c4e13b8 with --max-depth doubled and diff offPct against this run to isolate how much of the 19% is a truncation artifact vs a real lane mismatch · Re-classify the last 20 architect out-of-lane events using Relevant-Rooms-union membership instead of Originating-Terminal-only and measure what fraction resolve to in-lane · Seed the architect reef dictionary with a negative/exclusion list (blog/course/digestif/monologue vocabulary) and measure whether prewalkOverlap for future blog-tagged-architect commits drops toward 0, making the mismatch visible earlier in the pipeline rather than after a full walk

     Measured outcome (deterministic, not self-graded): room mean offPct 24→20 over 4 commits since · dictionary ablation saved 1141B

     A structurally-explained, correctly-content commit reading as 19%-over-threshold drift inflates this lane's apparent loss frequency with false positives, which is worse for premium than a real miss because it teaches the underwriting model to distrust a signal that is actually sound — restored the moment the Relevant-Rooms-union classifier ships and re-scores the 189 open architect events.

     Open the signed attestation →

149. 2026-08-31 · 📐 architect · offPct 31 · mistagged

     feat(hiring): the lightning rod is the determinism rot — one core, four twists, 96 across

     Why it left the lane: The commit's Originating-Terminal trailer is 📐 architect, but every touched file (docs/voice/ice-cream-not-vanilla-read.md, src/components/CrewBar.tsx, src/lib/crewHooks.mjs, tests/hiring/crew-bar-rotates.test.mjs) sits outside architect's owned surface (docs/strategy/, docs/03-product/, .workflow/rooms/*.html, .workflow/cognitive-dashboard.html, docs/01-business/strategy/, CLAUDE.md) and squarely inside a voice/copywriting + hiring-UI surface. On top of that, the new docs file and the crewHooks.mjs comment block literally quote the string 'CLAUDE.md' at least twice ('CLAUDE.md: QWEN IS BANNED FROM ACTIVE TASKS'; 'the canon is CLAUDE.md's DETERMINISM SLIP') — and CLAUDE.md is itself a literal entry in architect's owned-glob list. A gzip-NCD reef sensor scores on text content, so those two bare-string mentions register as architect-owned vocabulary even though no byte of CLAUDE.md was touched, seeding a small false-positive intent anchor that the walk then had to spread away from.

     The semantic pull: Two forces pulled in opposite directions: (1) the literal substring 'CLAUDE.md' appearing inside prose/comments (not as an edited file) created a false intent-lane seed toward architect's owned CLAUDE.md coordinate; (2) the overwhelming majority of real content mass — 'steelman', 'ghost-read', 'ice cream vs vanilla', 'crew hook', 'hiring' vocabulary, the CrewHook typedef rename (lede/body → hook/call), the CTA copy rewrites — is voice-craft and hiring-UI material with no home in architect's declared surface at all, so the walk lit 143 reality cells (vs 131 intent cells) largely outside the lane the false CLAUDE.md seed implied.

     Intent–reality gap: The commit message and Originating-Terminal trailer claim this is architect-room work, but the receipt measured 100% of the touched files landing on a voice/copywriting + hiring-UI surface with zero files in architect's declared globs — the claimed lane captured none of the actual diff.

     • Unify and merge the three independently-spec'd `room-drift-<room>-check.sh` variants named in ff09774e5 instead of letting a fourth (architect, navigator, laboratory, now this one) get re-diagnosed and re-dropped.
     • Add/confirm `.workflow/rooms/voice.html`'s owned-surface glob list includes `docs/voice/**`, `src/components/CrewBar.tsx`, `src/lib/crewHooks.mjs`, `tests/hiring/**` so this commit's real surface has a legitimate home.
     • Add a reef-seed filter that excludes a room-owned filename appearing only as a quoted string in prose/comments (not as an edited path) from that room's textual vocabulary bag.
     • Add a regression fixture: a synthetic commit with `Originating-Terminal: 📐 architect` but 100% of touched files under a different room's glob must classify to that other room, not architect — guards the exact defect measured here.
     • Cross-link this event's finding back into ff09774e5's ledger entry so the next room that independently re-derives the same fix finds it already scoped rather than re-diagnosing a fourth time.

     What we learned: A correct diagnosis sitting in the amnesia ledger is not self-executing — ff09774e5 named this exact mechanism 8 commits ago (Originating-Terminal-only lane classification contradicts CLAUDE.md's own sanctioned-divergence rule) and reported the fix as independently spec'd by three separate rooms, yet none of them shipped it, and this room's mean offPct is still climbing (188 out-of-lane events in 14 days). The structural fact this event adds: the reef sensor can be pulled into a false-positive lane seed by a BARE STRING MATCH on another room's owned path name appearing in prose (here, 'CLAUDE.md' quoted twice in comments/docs) even when zero bytes of that path were touched — meaning a diagnosis-without-a-shipped-guard doesn't just fail to fix recurrence, it compounds with an unrelated textual-quoting artifact to make the same class of event look worse than the underlying work actually is.

     The missing guard (the deliverable): A regression test asserting the drift-receipt's lane classifier ignores Originating-Terminal in isolation and instead resolves lane from Relevant-Rooms × owned-surface glob match — fixture: commit trailer `Originating-Terminal: 📐 architect` + `Relevant-Rooms: 📐 architect, 🎩 operator, 🎤 voice` with 100% of changed files under voice's declared globs must classify LANE=voice, not architect.

     Detector probe: signal attribution — gzip reef 70% · ballistic walk 30%. Right-sized, not shallow or saturated in a bad sense: litIntent=walksIntent and litReality=walksReality means every walk step is landing on real reef content on both sides (no wasted/dead steps), and the reality-vs-intent spread ratio (143/131 ≈ 1.09x) is modest rather than explosive. No depth adjustment indicated; the leverage is entirely upstream at the seed/reef stage. Experiments: Re-run this commit's placement with the reef seed restricted to touched-file paths only (strip inline-comment prose mentions of other files' names) and measure whether offPct drops below the 15 threshold. · Build a small corpus of commits that quote a CLAUDE.md-owned rule name in prose without touching CLAUDE.md itself, and measure false-positive lane-seed rate before/after excluding bare prose mentions from the reef vocabulary bag. · Ship the room-lane-classify.mjs glob-match step from ff09774e5 and re-run it against the room's last 188 out-of-lane events to measure what fraction were Originating-Terminal-only misclassifications (mistagged) vs genuine cross-lane work (structural/intentional) — this event's own classification implies the fraction is high.

     Measured outcome (deterministic, not self-graded): room mean offPct 24→19 over 10 commits since · dictionary ablation saved 984B

     An out-of-lane event that traces to a known-but-unshipped classifier bug plus a one-off textual false-seed is a bookkeeping loss, not a control failure — premium is restored by merging the already-spec'd glob-based lane classifier rather than by treating each recurrence as a fresh underwriting event.

     Open the signed attestation →

150. 2026-08-31 · 🎤 voice · offPct 29 · structural

     fix(milestones): the progression had no legend — four cold readers all stopped on the first row

     Why it left the lane: The diff's real work is scripts/milestones/progression.mjs (new file), tests/milestones/progression-carries-legend.test.mjs (new file), and edits to data/milestones.json, data/milestones-progress.json, data/milestones-next.json, data/milestones-market-contact.json. None of those paths match any glob in voice's declared owned surface (books/tesseract/chapters/*.md, src/content/blog/*.mdx, docs/05-content/blog/scratchpad/, decentralizedtalk/slideData.ts, src/app/decentralized*, src/app/6needs/, scripts/voice-*-check.sh) — the room recognizes prose and slide content, not board-data JSON or a derivation script. No region narrative is cached for this sha, so I anchor on the telemetry instead: prewalkOverlap=6 against litIntent=110 (about 5%) means the reef's seed match between the commit's claimed room and its actual files was already thin before any walk ran; the walk then lit 128 reality cells against 110 intent cells, and 29% of that reality mass fell outside the intent lane — nearly double the 15% threshold. The pull isn't a stray keyword, it's a category mismatch: 'progression.mjs', 'milestones.json', 'axes', 'shell predicates' are board-tooling vocabulary with no home in voice's glob list.

     The semantic pull: Terms doing the dragging: 'scripts/milestones/progression.mjs', 'data/milestones.json', 'axes', 'shell predicates', 'tests/milestones/progression-carries-legend.test.mjs' — all board/data-pipeline vocabulary, zero overlap with voice's blog/book/slide glob vocabulary. No cached encircled coords exist for this sha to cite directly.

     Intent–reality gap: The commit message claims a reader-legibility fix — four cold readers, a missing glossary, prose that stopped a general counsel and an allocator — which reads as voice's core competency (READER POV FIRST). The receipt measured the opposite: the actual mass landed in a new derivation script and four rewritten JSON ledgers, a coordinate voice's owned surface does not cover.

     • Add `scripts/milestones/**`, `data/milestones*.json`, `tests/milestones/**` to the 🎩 operator room's owned-surface glob (the `.workflow/rooms/*.html` file matching operator)
     • Add a `Lane-Room:` trailer to CLAUDE.md's COMMIT MESSAGE STRUCTURE, computed from the top owned-surface match rather than the terminal the operator happened to be sitting in
     • Backfill-audit voice's other 97 out-of-lane events from the last 14 days for the same Originating-Terminal-vs-Relevant-Rooms mismatch pattern before treating each as a distinct drift class
     • Write tests/pmu-simulator/lane-matches-relevant-rooms.test.mjs asserting the receipt lane equals Relevant-Rooms[0] when it diverges from Originating-Terminal
     • Re-run the receipt for 91b45e7e9 keyed off 🎩 operator and confirm offPct drops under the 15 threshold

     What we learned: The receipt keys drift-lane off Originating-Terminal — which terminal was physically open — but the commit's own Relevant-Rooms trailer already computes the correct room via owned-surface glob matching, and for milestones-tooling commits dictated from the voice terminal the two disagree systematically (98 events in 14 days for this room alone). The detector never consults an answer the commit already carries in its own trailer.

     The missing guard (the deliverable): tests/pmu-simulator/lane-matches-relevant-rooms.test.mjs — for a commit whose Relevant-Rooms[0] differs from Originating-Terminal, assert the receipt's scored lane is Relevant-Rooms[0] (or that a MISMATCH is surfaced), never silently the physically-open terminal.

     Detector probe: signal attribution — gzip reef 60% · ballistic walk 40%. Right-sized, not shallow or saturated: walksIntent equals litIntent (110=110) and walksReality equals litReality (128=128), meaning every walk step produced a uniquely lit cell with no wasted duplicate walks and no truncation (walkTruncated=false). The wider reality spread reflects a legitimately broader file footprint (4 JSON files + 1 script + 1 test) rather than a runaway walk. Adjustment: don't change depth — widen the INTENT seed to include the claiming room's own owned-surface glob vocabulary alongside the commit message text, so intent's walk has a comparably rich seed to reality's. Experiments: Recompute this commit's receipt twice — once keyed by Originating-Terminal, once by Relevant-Rooms[0] — across the last 30 voice out-of-lane events and measure whether Relevant-Rooms[0] consistently produces lower offPct · Feed each room's Owned file/dir surface globs directly into the intent-side reef seed (not just the commit message) and re-measure prewalkOverlap on a held-out set of the 98 historical voice events · Ablate: add scripts/milestones/** and data/milestones*.json to operator's owned surface, recompute this sha's receipt, and confirm offPct drops toward or under 15

     Measured outcome (deterministic, not self-graded): room mean offPct 18→6 over 4 commits since · dictionary ablation saved 1063B

     98 out-of-lane events in 14 days for a room whose lane key is the physically-open terminal rather than its file-relevance match inflates voice's apparent loss ratio with drift that isn't its own risk and understates operator's/board-tooling's real exposure; premium is restored by re-keying the receipt to Relevant-Rooms[0] and closing the missing owned-surface glob for milestones tooling.

     Open the signed attestation →