+Viewing public record · signed out
One holder opened this record to anyone. Every number is read off their backup; the links below say what this is and how to read it without us.
Spec adherence
You are reading one holder's record of what their AI agents ran, opened to anyone by that holder — for the deployer, the underwriter or broker, and counsel.
+ThetaDriven · Patent pendingmanifest b1d9927f3962… · taken 2026-10-14 17:32 UTC · 0 days old
+Recompute it yourselfone command, no call to us
curl -X POST https://thetadriven.com/api/notary/query -H 'content-type: application/json' -d '{"op":"retrieve","car_sha":"b1d9927f39627458c1fae14720166bc6df6c1394791bccb62d68510b4cab2c28","at":"<now, ISO>","nonce":"<random bytes, hex>","caller":{"pubkey":"<your ed25519 public key, hex>","sig":"<your signature over the query bytes, hex>"}}' (the witnessed copy of this archive, signed and paid under your own key; the archive hash itself is recomputed by its holder, or a carrier they grant: npx thetacog-mcp verify b1d9927f39627458c1fae14720166bc6df6c1394791bccb62d68510b4cab2c28)+One vocabulary — every term on this page, read the same by every desk
Every term below reads the same for a Chief Risk Officer, a Staff Engineer and a Reinsurance Underwriter — hover a term for how each desk reads it.
- rung — one of the three comparison bands a tick is read against — rung 1 (the whole call vs a 1M window), rung 2 (the steering payload beside the spec it steers by — a size the injector chooses, not the bill), rung 3 (a prepared run vs a 1M window).
- lane — where a placed row landed against the spec declared before it ran: in lane (the row matches the declared spec) or off lane (it does not).
- break — a tick that fell below 99% on rung 2 or rung 3, counted only over the measured population, with the (rung, cause) pair that produced it.
- priced row — a row the tape carries a cost_usd for; a row with no cost reads UNMEASURED and is never counted as $0.
- manifest — the sha that names one backed-up archive — the single identity every reading on the page is read off of.
- UNMEASURED — the word a field reads when no record answers it — never a zero, never a guess: an absence with its own stated reason.
+Who this page is forthe responsible AI user who reports the process
This page is for the responsible AI user who reports the process. Every number below is read off what you backed up and what was witnessed, never a guess.
+Sufficient forwhere each row landed against the spec declared before it ran
What this page is sufficient for: where each row landed against the spec declared before it ran. What it is not: whether the work was good, which is undecidable (Rice, 1953).
+Tokens in one task — the tail an underwriter prices45 % less for the same task mix · p 0.013 · n 663
Each curve is the share of tasks that used more than a given number of tokens, log-log: 36 tasks run against a scope declared before they started, 627 without one, 29 Aug – 13 Oct 2026. The typical task is the median; the price of cover is set by the rare one at the far right, where a straight fall is a fat tail.
scope declared before the task · n 36 no declared scope · n 627 grey band: the null
Tokens in one task, same task mix45 % fewer with a declared scopen=663
How often one task uses more than a given number of tokens: 36 tasks run against a scope declared before they started, 627 without one, 29 Aug – 13 Oct 2026. The typical task is the median; the price of cover is set by the rare one at the far right.
One task is a prompt to the change it committed, or one unattended run of a written spec row to the commit it graded; its tokens are the model's input plus cache reads and writes, summed over every model call in it. The axes are log-log: a straight fall is a fat tail.
That declaring a scope caused the gap: the declared tasks were not randomised, and the task mix differs between the two arms.
What any single future task will cost — this is the record's own tail over 29 Aug – 13 Oct, not a forecast.
Recompute it yourself: node scripts/vna/token-tail.mjs --json.
Re-run the snapshot: if the permutation p (now 0.013) rises above 0.05, or the declared worst-5 % share is no longer below the undeclared one, the separation is gone.
- For the same task mix, the declared tasks used 45 % less in total (p 0.013).
- The worst 5 % of undeclared tasks carry 39 % of all their tokens, against 14 % declared.
- Worst task: 15.4M declared, 310.4M undeclared.
Computed when this page was requested, from the task record inside this backup (b1d9927f3962…), never from a committed snapshot. Recompute: node scripts/vna/token-tail.mjs --json.
Receipt written from the record ; this curve drawn from it .
+What this reading lets us promise1 promised · 4 not promised · 4 next
What this reading lets us promise, what it does not, and what to run next — each line is its claim; the evidence, the commands and the wedge each would license sit behind the + below.
PROMISED · 1 — measured and admissible on this backup
- where each unit landed against the spec declared before it ran, re-runnable from the commit
NOT PROMISED · 4 — unmeasured, underpowered, measured and lost, or undecidable
- a prepared spec/snowball reaches a green guard more often than chain-of-thought on the same rows
- the token bill under STEER is lower than PLAIN at the same complexity decile
- the work is correct
- which fast-dollar wedge survives its falsifier
NEXT · 4 — one step per open reading, and the row that measures it
- measure: a prepared spec/snowball reaches a green guard more often than chain-of-thought on the same rows — C563a
- measure: the token bill under STEER is lower than PLAIN at the same complexity decile — C519
- read the outcome the actor did not author: the row's guard, red at the parent and green at HEAD — C444
- measure: which fast-dollar wedge survives its falsifier — C563c
+The evidence behind each lineevery claim's reading, every command, and the wedge each would license
PROMISED — measured and admissible on this backup
- where each unit landed against the spec declared before it ran, re-runnable from the commit — 546 placements on this backup (518 in lane, 28 off lane); recompute with npx thetacog-mcp verify b1d9927f3962…
NOT PROMISED — unmeasured, underpowered, measured and lost, or undecidable
- a prepared spec/snowball reaches a green guard more often than chain-of-thought on the same rows — UNMEASURED — C563a has not written its reading — /var/task/data/vna/spec-vs-cot.json is not on this record
- the token bill under STEER is lower than PLAIN at the same complexity decile — median 6.2 % vs PLAIN, range [-467.2 %, 81.1 %], 26 of 45 cheaper, sign-test z 1.04 — inside the noise (z under 2), not admissible
- the work is correct — undecidable (Rice, 1953) — no reading on this page can carry it; the outcome a reader can check is a guard the actor did not author
- which fast-dollar wedge survives its falsifier — UNMEASURED — /var/task/data/strategy/wedge-falsifiers.json is not on this record; its wedges read as drafts to run, sending is the holder's act
NEXT — one step per open reading: the row and the command that measure it, and the wedge it would license
- measure: a prepared spec/snowball reaches a green guard more often than chain-of-thought on the same rows —
C563a·node scripts/vna/spec-vs-cot.mjswedge agent-token-drift-audit: UNMEASURED — /var/task/data/strategy/wedge-falsifiers.json is not on this record; its falsifier reads as a draft to run, sending is the holder's act - measure: the token bill under STEER is lower than PLAIN at the same complexity decile —
C519·node scripts/vna/steer-ab-tokens.mjs --jsonlicenses no wedge on its own - read the outcome the actor did not author: the row's guard, red at the parent and green at HEAD —
C444·node scripts/vna/steer.mjs verify <row> --since <head-at-dispatch>licenses no wedge on its own - measure: which fast-dollar wedge survives its falsifier —
C563c·node --test tests/ops/c563c-wedge-falsifiers.test.mjslicenses no wedge on its own
spec-prep cost: 222,284,426 steering tokens over 671 ticks — shown, never weighed against adherence; adherence is read off the placement, not off what the spec cost to write
rung 2 best case, rare: 99.990 % (≈ 10,509:1) over 671 ticks, against a declared 1M-window counterfactual — the steering payload's share of the spec corpus, a size the injector chooses, never proof of correctness; the bill vs PLAIN beside it: median 6.2 % vs PLAIN, range [-467.2 %, 81.1 %], 26 of 45 cheaper, sign-test z 1.04 — inside the noise (z under 2), not admissible
Sufficient for: which claims this backup's readings carry today, and what would move each open one. Not sufficient for: whether the work was good (Rice), or whether a wedge is a good business.
+Your two workloads, as tapes3920 rows · $108.2285 priced
Every backed-up row is one mark, in time order across both tapes. Drag the slider: each row shows its published encircled panel, its address, and the spend to it.
showing the newest 300 of 3920 rows — the full tape
+Why the tape pauses8 pauses longer than the gap threshold
- no rows · 23 h · 2026-09-30 19:08Z → 2026-10-01 18:05Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 7.8 h · 2026-10-01 22:16Z → 2026-10-02 06:01Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 82.1 h · 2026-10-02 06:19Z → 2026-10-05 16:23Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 66.5 h · 2026-10-05 16:46Z → 2026-10-08 11:16Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 9 h · 2026-10-08 14:17Z → 2026-10-08 23:20Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 23.4 h · 2026-10-09 03:52Z → 2026-10-10 03:14Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 13.8 h · 2026-10-10 03:21Z → 2026-10-10 17:08Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
- no rows · 47.3 h · 2026-10-10 20:58Z → 2026-10-12 20:13Z · UNMEASURED — the backup carries no dated commit log beside the tape, so whether commits landed in this window is not read here; its panel home lists 650 published commits not on the tape, undated, so not placed in this window
■ in lane · ■ off lane · ■ UNMEASURED (U) — under each lane: its rows and rows per hour, then how many landed in lane, off lane, UNMEASURED
518 in · 28 off · 1297 U
0 in · 0 off · 76 UUNMEASURED · n 0 in the rows shown · 76 on the full tape
0 in · 0 off · 2000 U
0 in · 0 off · 1 UUNMEASURED · n 0 in the rows shown · 1 on the full tape
witness lag UNMEASURED — 0 of 2077 deployed rows carry a received_at at or after their closed_at (2077 carry no received_at, 0 flagged witnessed) …
Witness lagUNMEASURED no row carries both timesn=0
witness lag UNMEASURED — 0 of 2077 deployed rows carry a received_at at or after their closed_at (2077 carry no received_at, 0 flagged witnessed) — the notary's received_at lives on the witness ledger until a witness receipt is folded onto the row
The median, with [min, max], of received_at − closed_at over the deployed rows that carry both times; a row received before it closed is clock skew and is not counted.
That the lag is gone — this line measures it and removes nothing.
A row with no received_at was never witnessed: its lag is UNMEASURED, never 0, and it is not in the median.
Send each deployed row's receipt to the notary; every row that comes back with a received_at joins the median.
Recompute it from the backed-up rows: 2077 deployed, 0 with both times, 2077 without.
no priced row in these 300 rows · $108.2285 spent before them
■ in lane · ■ off lane · ■ UNMEASURED
- sha
- e732cfdecf
- address
- UNMEASURED
- ShortLex name
- UNMEASURED — the row carries no address
- lane word
- UNMEASURED
- drift
- UNMEASURED
- off-lane
- UNMEASURED
- model · cost
- UNMEASURED — a CODING row carries no model or cost_usd; the spend is read off the DEPLOYED LLM rows
- spend to here
- $108.2285 over 76 priced rows; 3844 unpriced rows counted as UNMEASURED
CODING 518 in lane · 28 off lane · 1297 UNMEASURED · 277 breaks below 99 % · top cause: leaf stale or brief unresolved — the pay… · the full tapes
+The tail of your token spendquerying the tasks your backup carries
Querying the tasks your backup carries…
+Token readings per tick2 of 672 ticks crossed 99 % vs the 1M counterfactual
Token readings per tick — the bill vs PLAIN (measured), against a 1M-context run (counterfactual), and the steering payload's footprint (rung 2)
Every chart below draws all 672 ticks, each older tick a dot; the rung-2 chart carries the links (links showing the newest 60 of 672 ticks). Link every tick
2 of 672 ticks crossed 99 % vs the 1M counterfactual · 672 ticks (45 STEER · 627 PLAIN).
+How to read these chartsthe legend, the ladder, every break and its cause
Every tick is one point, in time order: one prompt to the commit it produced. ● chat tick · ◆ headless-runner / claude -p tick · one line through the ticks in time order, broken wherever a tick is UNMEASURED · teal above 99 % (the payoff), violet below · the axis reads in nines (90 · 99 · 99.9 · 99.99) · hollow: the decile's baseline rests on under 5 PLAIN ticks · the line: 99 % · two charts: REGULAR (vs PLAIN, MEASURED) and WORST CASE (vs 1M, COUNTERFACTUAL). Drawn from the series you backed up, not from this machine.
2 of 672 ticks crossed 99 % vs the 1M counterfactual (0 of them /steer; 0 ticks UNMEASURED — no PLAIN tick in their decile) · 0 crossed 99 % vs PLAIN · 672 ticks (45 STEER · 627 PLAIN).
The ladder: three charts side by side in severity order, each exact and re-runnable on the record; rungs 2 and 3 are configuration health — a tick below the line is detected with its likely cause, never a gate
configuration health: rung 2 394 of 671 ticks above 99 % (58.7 %) · rung 3 36 of 36 (100.00 %) · rung 1 0 of 672 — 99 % on rung 1 would be news. Where each rung stands against the 99 % claim: rung 1 the best 20 % of ticks reach 86.9 % — 12.1 points below the 99.00 % claim · rung 2 the best 20 % of ticks reach 99.85 % — 0.85 points above the 99.00 % claim · rung 3 the best 20 % of ticks reach 99.71 % — 0.71 points above the 99.00 % claim. Per day: 2026-09-25 r2 1/1 · 2026-09-27 r2 8/9 · 2026-09-28 r2 24/24 · 2026-09-29 r2 33/36 · 2026-09-30 r2 3/3 · 2026-10-01 r2 12/12 · 2026-10-10 r2 25/35 · 2026-10-11 r2 3/10 · 2026-10-12 r2 8/10 · 2026-10-13 r2 5/7. Detected (277): 2026-10-11 c722f9dc2d rung2 97.6 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-11 37e3fb6426 rung2 97.4 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-11 c9827c9802 rung2 97.3 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-11 6168f5389b rung2 97.2 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-12 30b8efa05e rung2 96.9 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-12 5b3435ff2b rung2 96.7 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-13 442b5bbe40 rung2 98.7 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for · 2026-10-13 4074318d1f rung2 98.3 % — leaf stale or brief unresolved — the payload is large for the spec it stands in for — never a gate.
277 breaks below 99 % (rung 2: 277, rung 3: 0) — ranked by cause: rung 2 × 225 — leaf stale or brief unresolved — the payload is large for the spec it stands in for · rung 2 × 52 — snowball not served — no hook payload live at any call of the tick — the promise is that every break is detected, placed, priced and dispatched, never that a break cannot happen.
The operator's prior for rung 3: we get between 70 and 130 X improvement (98.6 %–99.23 %) — the operator, 2026-10-01, verbatim (the C519 row) — a stated prior, unreceipted: no measurement row on the record carries it; drawn on rung 3 as the reference band
UNMEASURED — no panel published at https://thetadriven.com/, and none carried in this backup for d6a50a41fe — panels 762 of 1716 tape commits · 1158 in the repository · the backup carries the newest 90 (cap 200 panels / 768 KB) · publish it: node scripts/pmu/commit-triptych.mjs --commit d6a50a41fe --no-open --no-qwen, commit public/commit/d6a50a41fe/, then node scripts/vna/backup.mjs --upload --tapes
the encircled Δ panel for d6a50a41fe
2026-09-08 · d6a50a41fe · PLAIN/chair · 36 calls
- rung 1 · whole call vs 1M: 65.9 %
- rung 2 · steering payload vs spec: 0.0 % — DETECTED
- rung 3 · prepared run vs 1M: UNMEASURED
+This commit, read in fulleach reading with its cause, and the bill beside it
rung 1 · the whole call vs a 1M window: 65.9 %
rung 2 · the steering payload beside the spec it steers by, not the bill: 0.0 % — DETECTED: snowball not served — no hook payload live at any call of the tick
rung 3 · a prepared run vs a 1M window: UNMEASURED — not a prepared -p run (a chat tick) — rung 3 is read off prepared runs only
On spec and a fraction of the tokens carried as steering — the same commit, both re-runnable. The bill is the other number: the bill vs PLAIN at the same complexity decile: median 6.2 % lower for /steer (n 45, range −467.2 % to +81.1 %, 26 of 45 cheaper, 0 above 99 %).
Click any point below to pair its panel; the panel is read from your backup's panels.json (sha-matched), the rungs from your tick record. The encircled Δ panel for d6a50a41fe is the one Rust pipeline's 144×144 as the holder published it, never redrawn here.
Whole call against a 1M window0 % of ticks above 99 %n=672
RUNG 1 · SEVERE — the whole call against a 1M window — what the whole call costs against a 1M window (the sceptic's question) · 1 − (input + cache_create + cache_read of the tick) ÷ (calls × 1,000,000) — every call charged its whole context, the window charged in full, the same weighting on both sides; a tick above 99 % here is news and is called out · the best 20 % of ticks reach 86.9 % — 12.1 points below the 99.00 % claim · 0 of 672 ticks at or over 99 % · 0 of 672 ticks cross 99 %
y: the formula on the left · the line: 99 % · teal above, violet below.
Whether the work in a tick was right — undecidable (Rice, 1953); this counts tokens against a declared baseline.
Every limit of this reading is listed once, under Every way to compute it.
Every point links to its commit and its encircled panel; recompute the series from your own tick record: node scripts/vna/steer-ab-tokens.mjs.
Re-run it on the backed-up series: the count above 99 % is recomputed from the record, never typed — 0 of 672 today.
Steering payload beside the spec58.7 % of ticks above 99 %n=671
RUNG 2 · STEERING PAYLOAD FOOTPRINT — the steering payload beside the spec it steers by — how small the steering payload is beside the spec corpus it stands in for — a size the injector chooses, never the bill (the engineer's question) · 1 − Σ over the tick's calls of (hook_additional_context tokens live at that call, since the last /clear or compaction) ÷ (calls × min(1,000,000, corpus ÷ 4)) — both sides the same quantity on one bytes÷4 ruler, so a per-token price cancels (C519-P-INVARIANT); no payload served at all reads 0 and is flagged · the best 20 % of ticks reach 99.85 % — 0.85 points above the 99.00 % claim · 394 of 671 ticks at or over 99 % · 394 of 671 ticks cross 99 %
Rung 2 reads the steering payload beside the spec it steers by (1 − payload ÷ corpus) — a size the injector chooses, not the bill. Beside it, the bill vs PLAIN at the same complexity decile: median 6.2 % lower for /steer (n 45, range −467.2 % to +81.1 %, 26 of 45 cheaper, 0 above 99 %). A larger ratio exists only in the rare best case, against a 1M-token counterfactual window, and nothing here rests on it; rung 2 is never offered as the bill or as proof the work is right.
y: the formula on the left · the line: 99 % · teal above, violet below.
Whether the work in a tick was right — undecidable (Rice, 1953); this counts tokens against a declared baseline.
Every limit of this reading is listed once, under Every way to compute it.
Every point links to its commit and its encircled panel; recompute the series from your own tick record: node scripts/vna/steer-ab-tokens.mjs.
Re-run it on the backed-up series: the count above 99 % is recomputed from the record, never typed — 394 of 671 today.
Prepared run against a 1M window100 % of ticks above 99 %n=36
RUNG 3 · PREPARED RUN — a fully specified -p brief against a 1M window — what a prepared, fully specified run costs against a 1M window (the question of what the engine does when it is healthy) · 1 − ((input_tokens + cache_creation_input_tokens) ÷ num_turns) ÷ 1,000,000 — MEASURED off the verdict row's usage off the runner's verdict row: the new tokens each call newly added (its brief, its tool results), not the cached prefix it re-reads (rung 1 carries that whole-call view); the brief's snowball cap is shown beside it as the DECLARED ceiling, never as the reading; a chat tick is not a prepared run and reads UNMEASURED here · all 36 ticks clear the 99 % claim — the lowest reaches 99.14 % · 36 of 36 ticks cross 99 %
y: the formula on the left · the line: 99 % · teal above, violet below.
Whether the work in a tick was right — undecidable (Rice, 1953); this counts tokens against a declared baseline.
Every limit of this reading is listed once, under Every way to compute it.
Every point links to its commit and its encircled panel; recompute the series from your own tick record: node scripts/vna/steer-ab-tokens.mjs.
Re-run it on the backed-up series: the count above 99 % is recomputed from the record, never typed — 36 of 36 today.
The bill against PLAIN0 % of ticks above 99 %n=672
REGULAR — vs the PLAIN tick measured at the same complexity decile · 0 of 672 ticks cross 99 %
y: 1 − actual ÷ the decile's PLAIN median actual · the line: 99 % · teal above, violet below.
Whether the work in a tick was right — undecidable (Rice, 1953); this counts tokens against a declared baseline.
Every limit of this reading is listed once, under Every way to compute it.
Every point links to its commit and its encircled panel; recompute the series from your own tick record: node scripts/vna/steer-ab-tokens.mjs.
Re-run it on the backed-up series: the count above 99 % is recomputed from the record, never typed — 0 of 672 today.
Worst case against a 1M window0.3 % of ticks above 99 %n=672
WORST CASE — vs a full 1,000,000-token window on every call a PLAIN tick of that decile takes · 2 of 672 ticks cross 99 %
y: 1 − actual ÷ (median PLAIN calls of the decile × 1,000,000) · the line: 99 % · teal above, violet below.
Whether the work in a tick was right — undecidable (Rice, 1953); this counts tokens against a declared baseline.
Every limit of this reading is listed once, under Every way to compute it.
Every point links to its commit and its encircled panel; recompute the series from your own tick record: node scripts/vna/steer-ab-tokens.mjs.
Re-run it on the backed-up series: the count above 99 % is recomputed from the record, never typed — 2 of 672 today.
+Every way to compute itthe formulas, the baselines, the prediction and the per-day counts
| every way to compute it | reading | formula |
|---|---|---|
| HEADLINE · steering-context savings · per tick, token-weighted over its calls | 394 of 672 cross 99 % | 1 − Σ over the tick's calls of (tokens of every hook_additional_context attachment still live in the session at that call, since the last /clear or compaction) ÷ (calls × min(1,000,000, bytes of SPEC-VNA-COCKPIT.md + SPEC-FROM-TREE.md + spec-tree.json at the tick's HEAD ÷ 4)) — both sides the same quantity, one bytes÷4 ruler, so any per-token price cancels |
| per tick · raw · vs 1M (worst case) | 2 of 672 cross 99 % | 1 − (input + cache_create + cache_read of the tick) ÷ (median calls of the decile's PLAIN ticks × 1,000,000) |
| per tick · raw · vs PLAIN (measured) | 0 of 672 cross 99 % | 1 − (input + cache_create + cache_read of the tick) ÷ (median of the same sum over the decile's PLAIN ticks) |
| per tick · price-weighted · vs 1M | 75 of 672 cross 99 % | 1 − (input + 1.25 × cache_create + 0.1 × cache_read of the tick) ÷ (median PLAIN calls × 1,000,000) — cache reads at their billed rate |
| per tick · price-weighted · vs PLAIN | 0 of 672 cross 99 % | 1 − (input + 1.25 × cache_create + 0.1 × cache_read of the tick) ÷ (median of the same weighted sum over the decile's PLAIN ticks) |
| per call · raw · vs 1M | 0 of 672 cross 99 % | 1 − (the tick's mean call: raw sum ÷ calls) ÷ 1,000,000 — one call against one full window |
| per call · raw · vs PLAIN | 0 of 672 cross 99 % | 1 − (the tick's mean call) ÷ (median mean call over the decile's PLAIN ticks) |
| per call · price-weighted · vs 1M | 7 of 672 cross 99 % | 1 − (the tick's mean weighted call) ÷ 1,000,000 |
| per call · price-weighted · vs PLAIN | 0 of 672 cross 99 % | 1 − (the tick's mean weighted call) ÷ (median mean weighted call over the decile's PLAIN ticks) |
| exact cost — API-billed claude -p ticks only | 0 API-billed tick(s) with an exact cost · 36 worker-reported | cost_usd as the API returned it on the runner's verdict row (cost_source api); a worker-reported figure (cost_source worker) is shown by its source and never as exact; a chat tick carries no billed cost |
Which ticks crossed 99 % on the worst case — 2
- 2026-09-08 · 3998b7ac4f · p65 · PLAIN/chair · 99.10 % · cost UNMEASURED (no billed cost on a chat tick)
- 2026-09-10 · 646ad28363 · p97 · PLAIN/chair · 99.47 % · cost UNMEASURED (no billed cost on a chat tick)
Prediction, not a promise: the operator expects /steer subagent and -p ticks to cross 99 % against the worst case several times a day. Today: NOT YET MET. Break it: a day of /steer -p ticks whose worst-case savings cross 0.99 at least 3 times, read off this chart. The lever: the per-call floor is Claude Code's own system prompt + CLAUDE.md + brief (≥ 60k tokens), so crossings need a worker that carries the brief alone (C517's local door, or a raw API worker) — a separate row, not assumed here.
Crossings per day (how often, never an average): 2026-09-13: 0 of 1 · 2026-09-14: 0 of 10 · 2026-09-15: 0 of 7 · 2026-09-16: 0 of 19 · 2026-09-17: 0 of 12 · 2026-09-18: 0 of 5 · 2026-09-19: 0 of 12 · 2026-09-20: 0 of 2 · 2026-09-22: 0 of 5 · 2026-09-23: 0 of 15 · 2026-09-24: 0 of 30 · 2026-09-25: 0 of 1 · 2026-09-27: 0 of 10 · 2026-09-28: 0 of 24 · 2026-09-29: 0 of 36 · 2026-09-30: 0 of 3 · 2026-10-01: 0 of 12 · 2026-10-10: 0 of 35 · 2026-10-11: 0 of 10 · 2026-10-12: 0 of 10 · 2026-10-13: 0 of 7
Per week: 2026-W35: 0 of 9 · 2026-W36: 0 of 198 · 2026-W37: 2 of 200 · 2026-W38: 0 of 67 · 2026-W39: 0 of 61 · 2026-W40: 0 of 75 · 2026-W41: 0 of 45 · 2026-W42: 0 of 17
EXPECTED, the 1M counterfactual: EXPECTED(tick) = median calls of the PLAIN ticks in the tick's complexity decile (C510's ruler, both labels ranked together) × 1,000,000 tokens — every call of a 1M-context run carrying the full window; a declared counterfactual, never a run we made
Savings: savings(tick) = 1 − actual ÷ EXPECTED; actual = the tick's input + cache_create + cache_read tokens (cache-weighted also shown: writes ×1.25, reads ×0.1); never averaged over a day or a week — a day or a week is only how many ticks crossed the line in it
WORST CASE baseline: COUNTERFACTUAL — EXPECTED(decile) = median PLAIN calls in the decile × 1,000,000; a declared formula, not a run we made
REGULAR baseline: MEASURED — the PLAIN median actual (input + cache_create + cache_read) of the tick's complexity decile
Share this reading: /backup/adherence?manifest=b1d9927f3962… — a carrier or CFO reads it without trusting this machine (the series is countersigned where it is held). Not sufficient for: vs 1M is arithmetic against a declared counterfactual, not a run we made · neither line says the work was good (Rice) · the task mix inside a decile is uncontrolled · a price (none pinned, KR40).
- CODING: 1843 rows, spend UNMEASURED
- job-36e68501 (DEPLOYED): 76 rows, $108.2285 over 76 priced
- job-55ddc699 (DEPLOYED): 2000 rows, spend UNMEASURED
- job-d0eac3ae (DEPLOYED): 1 row, spend UNMEASURED
+Actuarial Balance Sheetour side · 7 of 7 measured · your side · you type it
Our side is only what this backup measured; your side is what you type. Every cell names the record it was read off.
+Our sidemeasured on this backup · 7 of 7
OUR SIDE · measured on this backup — each cell with the record it was read off.
- Spend on the record: $108.23 over 76 priced rowssource: your backed-up tapes · cost_usd summed over priced rows (twoTapes spend)
- Rows priced: 76 of 3920 rows carry a cost; 3844 read UNMEASURED, not $0source: your backed-up tapes · rows with a numeric cost_usd
- Priced window: 2026-09-30 to 2026-10-12 (12.16 days)source: your backed-up tapes · first and last ts of the priced rows
- Recorded spend per month: $270.81 per 30.4375 days at the recorded ratesource: spend on the record ÷ 12.16 window days × 30.4375
- RUNG 1 · SEVERE: crossed 99 % on 0 of 672 measured ticks; 672 below, each with its causesource: your backed-up tick record · savingsReading → ladder.health
- RUNG 2 · STEERING PAYLOAD FOOTPRINT: crossed 99 % on 394 of 671 measured ticks; 277 below, each with its causesource: your backed-up tick record · savingsReading → ladder.health
- RUNG 3 · PREPARED RUN: crossed 99 % on 36 of 36 measured ticks; 0 below, each with its causesource: your backed-up tick record · savingsReading → ladder.health
+Your sideyou type it
YOUR SIDE · you type it — the monthly spend you state; this record never guesses it.
- UNMEASURED: you have not typed your monthly spendsource: you, typed on this page — never stored, never sent
- Ratio: UNMEASURED: your side: you have not typed your monthly spendsource: recorded spend per month ÷ your monthly spend
+What this sheet is sufficient forand what it is not
Sufficient for: the share of a spend you state that this record has detected, placed and priced.
Not sufficient for: what your spend is (you say it, we do not), whether any action was good (Rice), or any accounting value of the record.
+What your backup holds about how the code was written8 of 9 fields read · the rest UNMEASURED with the reason
- rows on the tape 1843
- commits 836
- turns 300
- in lane 518
- off lane 28
- off-lane share of placed rows 5%
- encircled panels published 776 of 836 commits
- spend UNMEASURED — the flight tape records no cost per commit or turn — spend is read on the DEPLOYED tab
- last seen 2026-10-14T17:32:18.812Z
+Per spec — the ledger130 specs on the record
Per spec: showing the newest 5 of 130 specs (by last seen). Every spec
+Spec 13b0c3b4c6dd2d09…coding beside deployed, one row each
Actions, admissible share, off-lane share, gaps, broken links, witness lag and last seen — per domain, read off the backed-up tape (CODING) and the witnessed receipts (DEPLOYED); an empty field reads UNMEASURED with its reason, never 0.
| domain | actions | admissible | off-lane | gaps | broken | witness lag | last seen |
|---|---|---|---|---|---|---|---|
| CODING | 2 | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | 2026-10-14T03:10:14.763Z |
| DEPLOYED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED |
+Spec 42f358a8425f3f95…coding beside deployed, one row each
Actions, admissible share, off-lane share, gaps, broken links, witness lag and last seen — per domain, read off the backed-up tape (CODING) and the witnessed receipts (DEPLOYED); an empty field reads UNMEASURED with its reason, never 0.
| domain | actions | admissible | off-lane | gaps | broken | witness lag | last seen |
|---|---|---|---|---|---|---|---|
| CODING | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED |
| DEPLOYED · your own tape, not witnessed | 2000 | UNMEASURED | UNMEASURED | 0 | UNMEASURED | UNMEASURED | 2026-10-14T17:20:52.778Z |
+Spec a5cb8c9d2768a15f…coding beside deployed, one row each
Actions, admissible share, off-lane share, gaps, broken links, witness lag and last seen — per domain, read off the backed-up tape (CODING) and the witnessed receipts (DEPLOYED); an empty field reads UNMEASURED with its reason, never 0.
| domain | actions | admissible | off-lane | gaps | broken | witness lag | last seen |
|---|---|---|---|---|---|---|---|
| CODING | 1 | 100% | 0% | UNMEASURED | UNMEASURED | UNMEASURED | 2026-10-14T16:44:48.818Z |
| DEPLOYED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED |
+Spec eccaf25f7b9b1dc0…coding beside deployed, one row each
Actions, admissible share, off-lane share, gaps, broken links, witness lag and last seen — per domain, read off the backed-up tape (CODING) and the witnessed receipts (DEPLOYED); an empty field reads UNMEASURED with its reason, never 0.
| domain | actions | admissible | off-lane | gaps | broken | witness lag | last seen |
|---|---|---|---|---|---|---|---|
| CODING | 1 | 100% | 0% | UNMEASURED | UNMEASURED | UNMEASURED | 2026-10-14T16:45:32.209Z |
| DEPLOYED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED |
+Spec f80d7d403bed2bc3…coding beside deployed, one row each
Actions, admissible share, off-lane share, gaps, broken links, witness lag and last seen — per domain, read off the backed-up tape (CODING) and the witnessed receipts (DEPLOYED); an empty field reads UNMEASURED with its reason, never 0.
| domain | actions | admissible | off-lane | gaps | broken | witness lag | last seen |
|---|---|---|---|---|---|---|---|
| CODING | 1 | 100% | 0% | UNMEASURED | UNMEASURED | UNMEASURED | 2026-10-14T02:40:55.086Z |
| DEPLOYED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED | UNMEASURED |
+What to do nextfour steps, the last one recomputes it yourself
- Pick who should see this below — we start with us; add an insurer or warrantor once they say yes to a logo slot.
- Turn on "shared by default" if this reading should answer its link for anyone who holds it, with no request on file.
- Email access below — to your deployer, your company, or the person you pay. Nothing is sent from this page; the email is yours to send, from your own mail client.
- Recompute it yourself: curl -X POST https://thetadriven.com/api/notary/query -H 'content-type: application/json' -d '{"op":"retrieve","car_sha":"b1d9927f39627458c1fae14720166bc6df6c1394791bccb62d68510b4cab2c28","at":"<now, ISO>","nonce":"<random bytes, hex>","caller":{"pubkey":"<your ed25519 public key, hex>","sig":"<your signature over the query bytes, hex>"}}' (the witnessed copy of this archive, signed and paid under your own key; the archive hash itself is recomputed by its holder, or a carrier they grant: npx thetacog-mcp verify b1d9927f39627458c1fae14720166bc6df6c1394791bccb62d68510b4cab2c28). You don't have to take our word for it.
+Email access to this pagea mailto from your own client, nothing sent from here
Send the read-only link to your deployer, your company, or the person you pay for this — whoever should see how the coding record is holding up.
+What this page is sufficient for, and what it is notwhere each placement lands; never whether the work was good
HOW MANY of your placements land where, per domain, per spec — read off your backed-up tape (CODING) and your witnessed receipts (DEPLOYED), and re-runnable from them.
Whether any of that work was good is undecidable (Rice, 1953); this page does not claim it and no reading here stands in for it.
Sufficient for: where it landed, re-runnable from the commit: same bytes, same hash. NOT sufficient for: whether the work was good, which is undecidable (Rice, 1953); nothing here claims it. Rice (1953): no program can decide, in general, whether another program is correct.
Recompute it yourself: curl -X POST https://thetadriven.com/api/notary/query -H 'content-type: application/json' -d '{"op":"retrieve","car_sha":"b1d9927f39627458c1fae14720166bc6df6c1394791bccb62d68510b4cab2c28","at":"<now, ISO>","nonce":"<random bytes, hex>","caller":{"pubkey":"<your ed25519 public key, hex>","sig":"<your signature over the query bytes, hex>"}}' (the witnessed copy of this archive, signed and paid under your own key; the archive hash itself is recomputed by its holder, or a carrier they grant: npx thetacog-mcp verify b1d9927f39627458c1fae14720166bc6df6c1394791bccb62d68510b4cab2c28). You don't have to take our word for it.