What this deployer's agents did, how large one task got, and what the priced part cost. Every figure is read from their own backup; none is typed in, and a missing one says UNMEASURED.
These are readings of an AI-assisted software development process. An agent answering patients or customers is a different book, with its own and much tighter bar; no number on this page is offered for that use.
Nobody has to decide whether the agent was right; an underwriter prices how large the work can get against a limit declared before it ran.
The sealINTACT
INTACT: the sha256 of the stored archive, recomputed from its bytes at this read, equals its name 64e05aeb2d77…. A later edit to any byte breaks the hash.
Where this sits in your forms
Before the numbers: most of these tasks ran with no scope declared, and the deployed calls read off lane. That is not a verdict on the work. It is the part a carrier prices, and declaring the scope first is the lever the deployer holds.
What to hold against a limit: the largest single task this record observed, and the tail of the tasks declared before they ran, both on the curve below. The limit and the rate stay your desk’s.
The market has already answered on paper: ISO/Verisk filed the generative-AI endorsements CG 40 47, CG 40 48 and CG 35 08, effective 1 January 2026, and those are CGL forms, not cyber forms. An exclusion is what a market writes over a loss it cannot count, and a book that reads agentic risk as outside cyber is carrying that exposure unpriced.
What to ask the insured for: the recompute pack (the archive and its manifest, named by its own sha256), the scope they write down before each agent task runs, and the endorsement wording once counsel clears it.
The tail below is counted in tokens. Only rows the model API billed carry a dollar price; every other row reads UNMEASURED, never $0. Pricing the risk stays your desk's act.
Observed, not yet shown to be caused: the declared tasks are the ones someone chose to scope, so part of the gap may be which tasks got scoped. Until the scope is assigned at random, read the gap as observed.
We measure where your agents’ work landed and attest it. We do not control or halt an agent; the halt is yours to wire. The countersign on these receipts is ours today. What you can rely on without us is the recompute: re-hash the archive and rerun the reading yourself.
Risk envelope
What this deployer's shared backup records, read for pricing. Each line is filled from the record or says UNMEASURED and why: detected, placed, priced, dispatched, and recomputable by you.
Rows on the shared backup, by domain: CODING 1883 · DEPLOYED 2095 — the holder's own tape; notary-witnessed rows are UNMEASURED here (read on the notary ledger)
Rung 3 calls above 99 %: DEPLOYED 95 of 95 · CODING 38 of 38
Rung 2 calls above 99 %: DEPLOYED 81 of 95 · CODING 407 of 674
Recompute it yourself: ask the deployer for the bundle and run npx thetacog-mcp verify <bundle url>. You don't have to take our word for it.
What a carrier still needs
What this reading is not sufficient for.
Whether any of this work was good: undecidable (Rice, 1953), and nothing here claims it.
Fault in law, or who pays: the attribution column is a reading of the recorded cause, never a verdict.
Outcomes in the world: the tape counts actions and where they landed, not losses; a loss history comes from the deployer and your own claims file.
A notary-witnessed chain and witness lag: the holder's own tape is not witnessed; those are read on the notary ledger.
Anything the agent did that was never written to the tape: a gap in a sequence is counted, never reconstructed.
A rate: nothing here is a filed rate or a price; pricing the risk stays your desk's act.
Specimen endorsement CY-2026-DET: drafted, with our counsel, not yet cleared. It will describe a crossing that was detected, placed and priced.
The trophy case · what this record lets you say, and how an underwriter checks each line · press + on any of them
+Tokens in a typical task3.7M tokens, the median taskn=675
HOW TO READ IT
LEFT BANK · deployer
My typical task runs on 3.7M tokens, over 675 tasks, and the whole curve is on this page.
RIGHT BANK · underwriter
The median of 675 per-task token counts, each the model's input plus cache reads and writes summed over the task's calls; recompute it from the tick record with node scripts/vna/token-tail.mjs --json.Read from: data/vna/steer-ab-tokens.ndjson in the archive, folded by savingsReading, sized by tailArms (C647c): the median task.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as the per-task survival curve (TaskTail, C650).
THE MITIGATION
LEFT BANK · deployer
an underwriter asks for the recompute pack (the Ask the holder box)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/token-tail.mjs --json. You don't have to take my word for it.
+Tasks run against a declared scope7.4% of tasksn=675
HOW TO READ IT
LEFT BANK · deployer
7.4 % of my 675 tasks ran against a scope I wrote down before they started.
RIGHT BANK · underwriter
Count the STEER rows among the 675 measured tasks in the tick record: 50 of 675. Recompute with node scripts/vna/token-tail.mjs --json.Read from: the same tick record: the share of measured tasks labelled STEER (a scope written before the task ran).
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as the running share, task by task, oldest first.
THE MITIGATION
LEFT BANK · deployer
the holder buys actions: notary queries and attested backups (/pricing)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/token-tail.mjs --json. You don't have to take my word for it.
+Worst task, declared and undeclared15.4M declared · 310.4 M undeclaredn=663
HOW TO READ IT
LEFT BANK · deployer
My largest declared task used 15.4M tokens; my largest undeclared one used 310.4M.
RIGHT BANK · underwriter
The maximum per arm over 663 tasks — the largest observed, never a limit on the next one. Recompute with node scripts/vna/token-tail.mjs --json.Read from: the same tick record: the largest task in each arm (tailReading max).
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as the two curve ends on the per-task survival curve.
THE MITIGATION
LEFT BANK · deployer
an underwriter asks for the recompute pack (the Ask the holder box)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/token-tail.mjs --json. You don't have to take my word for it.
+Bill carried by the worst 5 % of tasks34% undeclared · 14 % declaredn=625
HOW TO READ IT
LEFT BANK · deployer
The worst 5 % of my undeclared tasks carry 34 % of their tokens; declared, 14 %.
RIGHT BANK · underwriter
share5 per arm over 663 tasks. It is not monotone in the tail (C657), so it describes, it does not test; the tested figure is the share above a fixed size. Recompute with node scripts/vna/token-tail.mjs --json.Read from: the same tick record: share5 per arm (token-tail.mjs) — a description of concentration, not the tail test.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as two points, declared then undeclared.
THE MITIGATION
LEFT BANK · deployer
an underwriter asks for the recompute pack (the Ask the holder box)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/token-tail.mjs --json. You don't have to take my word for it.
+Steering payload beside the spec it steers by99.5% smaller than the specn=674
HOW TO READ IT
LEFT BANK · deployer
The steering my agents carry is 99.5 % smaller than the spec it steers by, on the median of 674 ticks.
RIGHT BANK · underwriter
Rung 2 per tick, 1 − Σ payload ÷ (calls × min(1,000,000, corpus ÷ 4)); a size the injector chooses, never the bill. Recompute with node scripts/vna/steer-ab-tokens.mjs --json.Read from: the same tick record: each tick's steering ledger reading (1 − payload ÷ the spec it stands in for), every tick that carries one (C647a).
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as one point per tick, oldest first.
THE MITIGATION
LEFT BANK · deployer
the holder buys actions: notary queries and attested backups (/pricing)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/steer-ab-tokens.mjs --json. You don't have to take my word for it.
+Tokens per unit of finished work1,013,102tokens per cogn=61
HOW TO READ IT
LEFT BANK · deployer
Each unit of finished work cost 1M tokens on the rolling median, over 61 minted cogs.
A cog is one minted unit of finished work: a commit whose guard failed on its parent and passes on it, closing or touching a spec row.
It is minted from the commit, never from the prompt: node scripts/vna/cog.mjs --since <sha> weighs the rows ticked, the basins touched and the cells lit (data/vna/cog-weights.json); a commit with no red witness is not minted.
The number is the rolling median of the newest 7 measured cogs; the line falling means each unit of finished work costs fewer tokens.
Fewer tokens per cog is health: the same finished work, bought with less.
RIGHT BANK · underwriter
A cog is minted from a commit whose guard failed on its parent and passes on it; tokens per cog against the band's peg. Recompute with node scripts/vna/cog.mjs --since <sha> && node scripts/vna/peg.mjs.
Each band is read against its peg: the fewest tokens per cog ever measured on a minted row of that band. The peg only falls.
n per band — band l: n=31, peg 192k, falling · band xl: n=26, peg 198k, falling · band m: n=3, peg 2.4M, below the floor of 10 · band s: n=1, peg 23M, below the floor of 10. A band under 10 rows draws points and no line.
Minted record: 2026-09-18 to 2026-09-19.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was right — undecidable (Rice, 1953). This counts tokens against minted work, never quality.
RIGHT BANK · underwriter
A dollar figure: no price per token is pinned here. Nor a trend in a band below its n floor.
THE MITIGATION
LEFT BANK · deployer
the holder buys actions: notary queries and attested backups (/pricing)
Recompute it from your own record: node scripts/vna/cog.mjs --since <sha> && node scripts/vna/peg.mjs.
RIGHT BANK · underwriter
Every point names its commit; the spend of record is data/vna/cog-spend.ndjson and the peg moves are data/vna/cog-peg.ndjson — a point whose tokens differ from that commit's usage rows falsifies it.
+Commits whose anatomy recomputes MATCHUNMEASUREDno anatomy rows in this backupn=0
Run node scripts/vna/anatomy.mjs --recompute --all --no-git on the archive: MATCH, MISMATCH or UNMEASURED per commit. A share is never printed off fewer than 10 rows.Read from: data/vna/anatomy-tape.ndjson in the archive (C648): the rows it carried; the MATCH share is computed where the bytes are.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as MATCH of n, once a recompute reports it.
THE MITIGATION
LEFT BANK · deployer
an underwriter asks for the recompute pack (the Ask the holder box)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/anatomy.mjs --recompute --all --no-git. You don't have to take my word for it.
+Rows with no licence stampUNMEASUREDcounted on the tape, not on this pagen=0
HOW TO READ IT
LEFT BANK · deployer
Every row I cannot vouch for is counted and shown grey, never hidden.
RIGHT BANK · underwriter
Count the rows whose licence reads UNLICENSED on the tape itself: node scripts/vna/outside-the-wording.mjs.Read from: the flight tape's licence stamp per row (C90 licenceOf; C647l's count) — the site's receipt rows do not carry the stamp.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as grey marks on the tape strip.
THE MITIGATION
LEFT BANK · deployer
a carrier claims a routing slot (data/underwriting/counterparties.json, an open slot until it says yes)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/outside-the-wording.mjs. You don't have to take my word for it.
+Largest drift of one commit3.5% at most · tolerance not declaredn=753
HOW TO READ IT
LEFT BANK · deployer
My largest drift on one commit was 3.496 %, over 753 placed commits.
RIGHT BANK · underwriter
The maximum of delta.driftPct over the 753 commit rows that carry one. A declared tolerance does not ride the tape yet, so it reads UNMEASURED. Recompute with node scripts/vna/flight-tape.mjs --tail 2000.Read from: the flight tape as backed up: delta.driftPct per commit row (the drift receipt).
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as drift per commit, in tape order.
THE MITIGATION
LEFT BANK · deployer
an underwriter asks for the recompute pack (the Ask the holder box)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/flight-tape.mjs --tail 2000. You don't have to take my word for it.
+Semantic vega2.62advisoryn=200
HOW TO READ IT
LEFT BANK · deployer
My drift spreads at a vega of 2.62, over 200 series rows.
RIGHT BANK · underwriter
Advisory dispersion of the drift series, never a filed rate. Recompute with npx -y thetacog-mcp@latest vega-backtest.Read from: public/.well-known/semantic-vega.json in the archive (C648a): how widely the drift readings spread.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as the vega pill, apart from the tail.
THE MITIGATION
LEFT BANK · deployer
an underwriter asks for the recompute pack (the Ask the holder box)
RIGHT BANK · underwriter
Recompute it: npx -y thetacog-mcp@latest vega-backtest. You don't have to take my word for it.
+Hook latency, p50 and p95UNMEASUREDkept on the holder's machine, not in the backupn=0
Per-hook p50 and p95 off the hook timing ledger; recompute with node scripts/vna/hook-latency.mjs --hours 6.Read from: data/vna/hook-latency.json on the holder's machine — not carried in the backup.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as p50 and p95 per hook.
THE MITIGATION
LEFT BANK · deployer
the holder buys actions: notary queries and attested backups (/pricing)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/hook-latency.mjs --hours 6. You don't have to take my word for it.
+Backed up, and what it costs to hold2.8MB held · $0.000060 a monthn=1
HOW TO READ IT
LEFT BANK · deployer
This backup is 2.8 MB and costs $0.000060 a month to hold.
RIGHT BANK · underwriter
Receipt bytes × the store's published rate, read 2026-10-04. Recompute with node scripts/vna/backup-cost.mjs --snapshot public/backup/backup-cost.json.Read from: the backup receipt's bytes × public/backup/backup-cost.json rate.storage_usd_per_gb_month (C647j); coverage is C651a's walk-segment acknowledgement.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the work was good. That is undecidable (Rice, 1953), and no stat here claims it.
RIGHT BANK · underwriter
A loss figure or a rate. It is this holder's own record, drawn as the backup's size.
THE MITIGATION
LEFT BANK · deployer
the holder buys actions: notary queries and attested backups (/pricing)
RIGHT BANK · underwriter
Recompute it: node scripts/vna/backup-cost.mjs --snapshot public/backup/backup-cost.json. You don't have to take my word for it.
$168.30spent to here, on the tasks that record a cost
$4.43 per priced taskworst task $15.00
Over 38 of 675 backed-up tasks that record a billed cost (0 billed by the API, the rest as the worker reported). The other 637 are unpriced, never counted as $0.
+Tokens in one task — the tail an underwriter prices39 % less for the same task mix · p 0.014 · n 663
Each curve is the share of tasks that used more than a given number of tokens, log-log: 38 tasks run against a scope declared before they started, 625 without one, 29 Aug – 15 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 38 no declared scope · n 625 grey band: the null
+Tokens in one task, same task mix39% fewer with a declared scopen=663
HOW TO READ IT
LEFT BANK · deployer
How often one task uses more than a given number of tokens: 38 tasks run against a scope declared before they started, 625 without one, 29 Aug – 15 Oct 2026. The typical task is the median; the price of cover is set by the rare one at the far right.
RIGHT BANK · underwriter
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.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
That declaring a scope caused the gap: the declared tasks were not randomised, and the task mix differs between the two arms.
RIGHT BANK · underwriter
What any single future task will cost — this is the record's own tail over 29 Aug – 15 Oct, not a forecast.
THE MITIGATION
LEFT BANK · deployer
Recompute it yourself: node scripts/vna/token-tail.mjs --json over that tick record.
RIGHT BANK · underwriter
Re-run the snapshot: if the permutation p (now 0.014) 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 39 % less in total (p 0.014).
The worst 5 % of undeclared tasks carry 34 % of all their tokens, against 14 % declared.
Worst task: 15.4M declared, 310.4M undeclared.
Read from the tick record this backup carried (data/vna/steer-ab-tokens.ndjson, archive 64e05aeb2d77…), through the same tail function as the reference curve. Nothing here is ours. Recompute: node scripts/vna/token-tail.mjs --json over that tick record.
Receipt written from the record ; this curve drawn from it .
Vega · how much the drift itself swings · its own graph, not the tail
+Your semantic vega2.62advisory · seal UNMEASUREDn=200
HOW TO READ IT
LEFT BANK · deployer
How steady your agents' work has been. Each commit lands at a place in the declared scope; the vega is how widely those landings spread across the last 200 commits. Lower is steadier. The line shows every sealed commit, so you can see when it changed.
RIGHT BANK · underwriter
Each point is the sample spread of the drift readings over the 200 commits ending there, the same window the published print prices. The line runs across the sealed archive and the current window. The last point is the number on the print, recomputed here from the sealed rows.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether any commit was good. That is undecidable (Rice, 1953), and this number does not claim it. It says how far the work moved from where it was declared, not what the moves cost.
RIGHT BANK · underwriter
A filed rate or a loss figure: it is ADVISORY and raw, with the scope correction not yet applied. Nor does one print separate a change in behaviour from a slide of the window, which is why the whole line is drawn.
THE MITIGATION
LEFT BANK · deployer
Publish your own. The same two commands print your vega from your own tape, and 🔗 Countersign (backup) carries it to whoever reads your record.
RIGHT BANK · underwriter
Price from the line, never one print. Ask for the series and the seal and recompute both; a seal that reads BROKEN means a priced row changed outside the one writer, and the print is not evidence until it is re-sealed.
2.62 · ADVISORY · over 200 series rows, as the backup carried it
The recompute · you don't have to take my word for it
UNMEASURED — node scripts/vna/anatomy.mjs (the 🔗 backup carries the rows taped since the last one)
Sufficient for:what this holder's backup records: how large one of their tasks got, what the priced part cost, how widely the drift placement spread, and which commits it carries, each recomputable from the archive. NOT sufficient for:whether any of the work was good (undecidable, Rice 1953), a loss figure (tokens measure work, not loss), or anyone's agents but this holder's. Rice (1953): no program can decide, in general, whether another program is correct.
Backed up
15 Oct 2026
Tasks
675
Commits
879
Model calls
2,095
Tasks span
29 Aug 2026 – 15 Oct 2026
Share this page
Send this link to whoever you want to show: a broker, an underwriter, a peer. It opens read-only, with no account.
They see the per-task curve, the money, the vega and the panels from this backup, under a pseudonym.
They never see your name, your email, your prompts or your source code.
What your backup carries for a recompute
UNMEASURED — node scripts/vna/anatomy.mjs
The setting lives in data/vna/anatomy-setting.json. Tuple only keeps every byte of your code on your machine; tuple + spans lets an underwriter recompute without it.
What this backup costs to hold
2.7 MB held · $0.000060 a month to keep at $0.0213 per GB-month (Supabase Storage, price read 2026-10-04)
One story, three readings
Token efficiency, the fat tail and vega, read off this backup
+1 · Token efficiency99.5% smaller than the spec it steers byn=674
HOW TO READ IT
LEFT BANK · deployer
The steering your agents carry is small beside the spec it steers by: on the median tick it is 99.5 % smaller. Each point is one tick, oldest first.
RIGHT BANK · underwriter
Rung 2 per tick: 1 − Σ payload ÷ (calls × min(1,000,000, corpus ÷ 4)), over 674 ticks. Read from the tick record this backup carried.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
The bill. It is the size of what steers the work, chosen by the injector, not what the work cost.
RIGHT BANK · underwriter
A causal claim about the tail beside it. Whether this smallness thins the tail is what the coin-flip trials test.
THE MITIGATION
LEFT BANK · deployer
Keep the spec written down and the steering small; the ledger shows each tick that ran without it.
RIGHT BANK · underwriter
Recompute it from the tick record: node scripts/vna/steer-ab-tokens.mjs --json. You don't have to take my word for it.
+2 · The fat tail45.2% of undeclared tokens sit above 11.6M · declared 8.3 %n=663
HOW TO READ IT
LEFT BANK · deployer
How much of the bill sits in the largest tasks. Above one fixed size, 11.6M (the undeclared p90), undeclared tasks carry 45.2 % of their tokens; declared tasks, 8.3 %. The line is that share at the undeclared median, p75 and p90.
RIGHT BANK · underwriter
shareAbove at a fixed K, per arm (C657's monotone tail figure), n 38 declared and 625 undeclared, observed on this record. As a description only: the worst 5 % of undeclared tasks carry 34 % of their tokens, declared 14 %. Read from the same tick record as the curve above.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
A size for the next task. The largest one observed says nothing firm about the next one.
RIGHT BANK · underwriter
A randomised comparison: the declared tasks were not assigned by a coin, so route and scope may be confounded. The worst-5 % share is not monotone in the tail, so it describes, it does not test.
THE MITIGATION
LEFT BANK · deployer
Declare the scope before the task starts; the curve shows which side your largest tasks fall on.
RIGHT BANK · underwriter
Recompute shareAbove from the survival arrays: node scripts/vna/token-tail.mjs --json, then the claim graph (src/lib/backup/claim-graph.mjs).
+3 · Vega2.62how widely the drift spreadsn=200
HOW TO READ IT
LEFT BANK · deployer
How much the drift itself swings from commit to commit: the volatility of where the work lands, not of what it costs. This record carries the print, not the series behind it.
RIGHT BANK · underwriter
The sample spread of the drift readings over the 200 rows ending at each point (C648a), advisory. A separate graph from the tail, never merged with it. Read from the vega print this backup carried.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether any commit was good (Rice, 1953).
RIGHT BANK · underwriter
A filed rate, or a cause of the tail. Vega is dispersion of placement; the tail is tokens per task.
THE MITIGATION
LEFT BANK · deployer
Publish your own print and carry it in the backup.
RIGHT BANK · underwriter
Recompute it: npx -y thetacog-mcp@latest vega-backtest, and check the seal.
On this record we observe three things together: the steering ran 99.5 % smaller than the spec it steers by (median of 674 ticks), declared tasks put 8.3 % of their tokens above 11.6M against 45.2 % undeclared (n 38 and 625), and the drift spread at a vega of 2.62. Whether one moves the other is the open question the coin-flip trials are built to answer; nothing here claims it.
At a glance · press + to read any of them
+Largest single task310.4M tokensn=675
HOW TO READ IT
LEFT BANK · deployer
Every task your agents ran, smallest to largest, from the tick record this backup carried. The line runs low and flat, then climbs at the right; that climb is the tail. The number is the most tokens one task used.
RIGHT BANK · underwriter
The holder's per-task token counts, sorted (input plus cache reads and writes, summed over every model call in the task). The last point is the maximum, 310.4M. The curve beside it is drawn from the same tasks.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether the largest task was waste or worth it. Tokens measure work, not loss, and whether the work was good is undecidable (Rice, 1953).
RIGHT BANK · underwriter
A loss figure, or the size of the next task. It is the largest observed over n tasks; a larger one can arrive tomorrow.
THE MITIGATION
LEFT BANK · deployer
Declare the scope before the work starts. The curve splits your tasks into declared and undeclared, so you can see which side your largest tasks sit on.
RIGHT BANK · underwriter
Price from the curve, never one number. You don't have to take my word for it: ask the holder for the archive, check its sha256 against its name, and read the tick record inside it (data/vna/steer-ab-tokens.ndjson); the curve is a pure function of those rows.
+Spend to here, priced tasks168.3USDn=38
HOW TO READ IT
LEFT BANK · deployer
What the priced part of the backed-up work cost, added up task by task. The line is the running total, so a step is one expensive task. $4.43 per priced task on average; the worst one $15.00.
RIGHT BANK · underwriter
The sum of cost_usd over the carried tasks that record one (38 of 675). 0 of them are billed figures from the API; the rest are as the worker reported them. A task with no recorded cost is left out, never counted as $0.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Your whole bill. Chat tasks record no billed cost, so the total covers the priced, unattended runs only.
RIGHT BANK · underwriter
A loss or a premium base. It is spend on work done, read off the holder's own record; the unpriced tasks are named, not filled in.
THE MITIGATION
LEFT BANK · deployer
Run more of the work as unattended, written-down tasks: each one records its own cost, so the total covers more of what you spend.
RIGHT BANK · underwriter
Read the priced share before the total, and ask for the tick record to recompute it. You don't have to take my word for it: the sum is over rows anyone can count.
+Commits with a panel carried90of 879 commitsn=90
HOW TO READ IT
LEFT BANK · deployer
Each panel is the picture of where one commit landed against the scope declared before it ran. This backup carried these, so they stay readable even if your repository goes private.
RIGHT BANK · underwriter
The PNG bytes rode inside the archive and are served from the backup store, never fetched from the holder's repository at read time. The line counts them up, one per carried commit.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether a commit was good. A panel places the work; it does not grade it (Rice, 1953).
RIGHT BANK · underwriter
Commits without a panel. The count says how many carry one, and the rest are named by the difference, never assumed clean.
THE MITIGATION
LEFT BANK · deployer
Back up after each working day, so every commit's panel rides with the record you send.
RIGHT BANK · underwriter
Spot-check a panel against its commit: the placement is recomputable from the commit itself, so ask the holder for the sha and re-run it.
RIGHT BANK · the underwriter's lead graph
Their semantic vega, as the backup carried it
+Your semantic vega2.62advisory · seal UNMEASUREDn=200
HOW TO READ IT
LEFT BANK · deployer
How steady your agents' work has been. Each commit lands at a place in the declared scope; the vega is how widely those landings spread across the last 200 commits. Lower is steadier. The line shows every sealed commit, so you can see when it changed.
RIGHT BANK · underwriter
Each point is the sample spread of the drift readings over the 200 commits ending there, the same window the published print prices. The line runs across the sealed archive and the current window. The last point is the number on the print, recomputed here from the sealed rows.
WHAT IT IS NOT SUFFICIENT FOR
LEFT BANK · deployer
Whether any commit was good. That is undecidable (Rice, 1953), and this number does not claim it. It says how far the work moved from where it was declared, not what the moves cost.
RIGHT BANK · underwriter
A filed rate or a loss figure: it is ADVISORY and raw, with the scope correction not yet applied. Nor does one print separate a change in behaviour from a slide of the window, which is why the whole line is drawn.
THE MITIGATION
LEFT BANK · deployer
Publish your own. The same two commands print your vega from your own tape, and 🔗 Countersign (backup) carries it to whoever reads your record.
RIGHT BANK · underwriter
Price from the line, never one print. Ask for the series and the seal and recompute both; a seal that reads BROKEN means a priced row changed outside the one writer, and the print is not evidence until it is re-sealed.
The vega your backup carried: 2.62, state ADVISORY, over 200 series rows, published 2026-10-15T15:51:47.866Z. Yours, read from your archive; nobody else's number is drawn here.
seal UNMEASURED — the backup carries the print, not the sealed ledger it was computed over
You don't have to take my word for it. Recompute it yourself:
The deployer shared this reading of their own backed-up record so you can price it. Their name and email stay with them; this page is keyed by a pseudonym. Every field is filled from the record or says UNMEASURED and why. None of it is a zero standing in for an absence.
Not sufficient for: how often an action failed in the world — the tape counts actions, not outcomes; whether any of this work was good — undecidable (Rice, 1953).
Off-lane share among placed actions
CODING 5% (26 of 534) · DEPLOYED 100% (87 of 87)
Not sufficient for: whether an off-lane action was harmful — only that it left the lane its owner declared; whether any of this work was good — undecidable (Rice, 1953).
Gaps in the deployed sequence
0 gaps across 3 jobs — a missing seq is itself a countable displacement
Not sufficient for: what the missing actions were — a gap is counted, never reconstructed; whether any of this work was good — undecidable (Rice, 1953).
Broken links
UNMEASURED — a link check needs the IntentGuard chain over witnessed receipts — read on the notary ledger (C495a), not on the holder's own backed-up tape
Not sufficient for: whether a break was malicious; whether any of this work was good — undecidable (Rice, 1953).
Regressed ticks
UNMEASURED — the backed-up tape rows carry no spec-tick history to compare against
Not sufficient for: why a declared row came back open; whether any of this work was good — undecidable (Rice, 1953).
Chain verdict
UNMEASURED — chain_verified and where it breaks are read on the notary ledger (C495a), not on the holder's own backed-up tape
Not sufficient for: anything about actions never sent to the notary; whether any of this work was good — undecidable (Rice, 1953).
Witness lag
UNMEASURED — the holder's own tape is not witnessed — witness lag is read on the notary ledger (C495a), not on the holder's own backed-up tape
Not sufficient for: what happened between action and witness; whether any of this work was good — undecidable (Rice, 1953).
Spend
$137.7642 over 94 priced rows · DEPLOYED $137.7642 (2001 of 2095 rows unpriced) · CODING UNMEASURED (the flight tape records no cost per commit or turn) — an unpriced row is UNMEASURED, never $0
Not sufficient for: any spend the tape did not record; whether any of this work was good — undecidable (Rice, 1953).
Semantic vega
vega 2.62 · state ADVISORY · 200 series rows · published 2026-10-15T15:51:47.866Z · recompute: npx -y thetacog-mcp@latest vega-backtest && npx -y thetacog-mcp@latest advisory-premium --series attest-out/backtest-series.ndjson
Not sufficient for: a rate — a vega is dispersion, advisory, never a filed rate; whether any of this work was good — undecidable (Rice, 1953).
Not sufficient for: this deployer in particular — they are the same for every deployer; whether any of this work was good — undecidable (Rice, 1953).
Encircled panels
UNMEASURED — no panel was copied at share time — the backup names no panel home, or none answered
Not sufficient for: whether the work was good; whether any of this work was good — undecidable (Rice, 1953).
Where this landed is re-runnable from the deployer's record. Whether the work was good is undecidable (Rice, 1953), and nothing here claims it. You don't have to take our word for any of it: run curl -X POST https://thetadriven.com/api/notary/query -H 'content-type: application/json' -d '{"op":"retrieve","car_sha":"64e05aeb2d770524ce32f1230663f10aa2a4b76374ea0ca309e5212aabece8c3","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 64e05aeb2d770524ce32f1230663f10aa2a4b76374ea0ca309e5212aabece8c3).
New to these backups? How to read one without us: what arrives, the free recompute, the paid query, where the keys are, and what it is not sufficient for.
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.
