On July 29, 2026, I counted 3,328 execution receipts in OrgX's own work record and one acceptance row. The acceptance table had only existed since July 25. Before that, there had been nowhere to record the ruling.
I had months of agent activity and verdicts. I had much less evidence of what a person accepted afterward. Those historical counts are the starting point of this essay, not a current adoption statistic or a fresh database audit.
A behavior verdict answers a narrower question
Behavior specs, introduced by Braintrust and Basis, describe how an agent should work. Their rendered-deck example asks a presentation agent to render, inspect, fix problems, and render again before claiming validation. A judge then assesses the relevant trajectory against the spec.
That makes an expectation explicit and testable. It also leaves another question open: did following that expectation produce work worth accepting?
Here is an illustrative failure. The agent inspects one version of a deck, then delivers a different version. A behavior judgment may establish that an inspection occurred. The reviewer still needs to know whether it covered the delivered artifact, whether the agent had authority to deliver it, and whether the deck answered the request.
The illustrative deck case separates process compliance from review of the delivered version.
A process check and the owner's judgment can disagree without either being a meaningless record. The mistake is to use one as a substitute for the other.
Keep the records distinct
| Record | Question it answers |
|---|---|
| Task | What was requested? |
| Behavior spec | How should the agent conduct the work? |
| Trace | What events occurred? |
| Evaluation | Did those events meet the standard? |
| Work receipt | What does the producer claim about this episode, with which evidence? |
| Work ledger | What do the accumulated episodes and rulings show? |
The diagram separates record types; linking them does not independently verify their claims.
The receipt is the point where those records can be linked. It can identify the evaluated standard, reference the trace and delivered artifact, and retain authority, cost, human intervention, and acceptance information. It need not embed an entire sensitive transcript to preserve those relationships.
That does not make every receipt independently verified. A producer can report a result or an acceptance claim; a consumer still has to inspect its source and authority.
The crosswalk has a useful edge case
The Agent Work Receipt v0.1 crosswalk maps behavior judgments into existing receipt fields. It identifies the judge, attaches the rationale as evidence, and pins the behavior specification by producer revision and digest.
A true or false verdict becomes a passed or failed check. NA needs more care. If the behavior was not applicable, the receipt marks the check skipped. If the evidence was insufficient or the behavior could not be judged, it marks the check inconclusive. A missing or failed judge result also stays inconclusive.
Those differences change what another system may infer. Skipping an irrelevant check is different from lacking the evidence to perform a necessary one. Converting both into a pass would discard the uncertainty at exactly the point another party needs it.
The published fixture shows the mapping. Its synthetic digests demonstrate format; they are not proof of retrievable source documents. Content integrity, signature validity, and signer trust remain separate checks.
Ask whether the standard helps
A ledger can connect the behavior judgment with what happened afterward: the owner's ruling, a correction, or an observed outcome. That gives a team material for revising the standard itself.
The proposed loop preserves later rulings. It does not show that the standard caused acceptance.
It does not establish causation. Work might be accepted for reasons the behavior spec never measured, or rejected because the request was wrong. A useful record preserves that possibility instead of awarding more autonomy merely because the visible process received a passing grade.
If you emit behavior verdicts, start with one NA result and follow it through the fixture. Check the reason, the artifact version, and any actual human ruling. That is a small enough case to reveal what the mapping loses before another system acts on it.