// Questions

What should an AI agent audit trail record?

Enough to re-check a claim later without trusting the agent that made it: the user's original request, each piece of evidence with its source and the exact target it observed, when it was observed, how far it reached (acknowledged or independently confirmed), and, for every approval, what the approver actually looked at.

Answered from GDF Research experiments by Sterling Morrison. Numbers link to the papers below.

Why each field earns its place

The original request: verifiers without it caught none of the wrong-target completions in our study, and verifiers with it caught most.

Source and target: seven checks that all read one source are one check. In simulation, treating them as seven pushed confidence to 0.99999994 on tasks that failed 6.1% of the time.

Time and world state: a fact observed before a code change isn't evidence about the code after it. A coding agent with an undated notes file trusted a stale "tests pass" in 5 of 5 runs.

Level of confirmation: a deploy API returning "success" is an acknowledgment, not proof the new version is serving traffic.

What each approver observed: when a verifier was told that seven earlier approvals rested only on the executor's message, it discounted them and accepted 2% of faulty work. Without that label, the same approvals pulled it to 40%.

Source papers

Related questions