# 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.

Author: Govna
Published: 2026-08-23
Updated: 2026-08-23

Evidence:
- [AUD-001](https://govna.io/solutions/security-leaders#AUD-001)
- [AUD-002](https://govna.io/solutions/security-leaders#AUD-002)

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](/solutions/security-leaders#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](/solutions/security-leaders#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](/glossary).