New One-call approvals are live in the console

Engineering essay

What an AI agent receipt proves—and what it does not

A receipt is evidence of what happened, not a guarantee of what will happen next.

Govna · Published · Updated

Give every agent a chain of command. A receipt is the artifact that lets someone outside the moment verify a piece of that chain: that a call’s terminal outcome correlates back to the audit intent committed for it. It is worth being precise about what that artifact does and does not establish.

Before dispatch is where the auditing actually happens, and the receipt is not that step — it is what survives it. Govna commits a durable audit intent before a connector call is dispatched, and if that commit cannot be made, the call is denied rather than allowed to proceed un-audited (AUD-001). The receipt does not perform that check; it is the terminal record of a call that already passed it.

What the receipt itself proves is narrower: that its terminal outcome correlates back to that pre-dispatch intent. Every dispatched call carries a correlation identifier threaded from intent through completion, so a given receipt can be tied to the specific intent it completed, rather than assembled after the fact from logs that might disagree with each other (AUD-002). AUD-001 is what stops a call from dispatching without a committed intent; AUD-002 is what lets a terminal receipt be correlated back to that intent by call identifier. That correlation is what the receipt proves on its own.

What a receipt does not prove is that the call was a good idea, that the upstream system did what the request asked, or that no future call will go wrong. A receipt is a record of what was authorized and dispatched — it is not a warranty on outcomes, and it is not a substitute for the policy decision that should have stopped a bad call before it reached the connector. Evidence proves what happened. It does not by itself decide what should have happened; that is what policy and human approval are for.

Treat a receipt as what it is: the durable, traceable record that a chain of command existed for one specific call. Read it that way, and it holds up. Read it as a broader guarantee than that, and you will be disappointed by exactly the gap it was never built to close. For how that call was decided in the first place, see fail-closed receipts.