AI agent log verification: how to check the record without trusting the agent
The short answer A log verifies when three separate questions all get answers. Consistency: does each record still match the hash recorded for it, so nothing inside the log was edited, dropped, or reordered. Completeness: does the log's head match an anchor held somewhere the log's owner cannot rewrite, so the tail was not cut and the whole file was not swapped. Coverage: for every step the agent…
Verification of an AI agent's log record requires three key checks: consistency, completeness, and coverage. Consistency involves recomputing a record's hash and comparing it to the published output_hash. Completeness examines whether the log's head matches an anchor held by the log owner to ensure the tail was not cut or the file swapped. Coverage verifies that for every step the agent claims, there is a corresponding record.
Most verification discussions stop at consistency. However, the second and third checks are crucial. Per-call receipts minted by the execution layer at runtime provide transparency. Each receipt is resolvable at its public URL and recomputable by anyone without an account. The verifier reports failure states explicitly, such as "verification failed," "rejected: unresolvable record," or "absent," rather than hiding them.
For AER-1 receipts, take the canonical bytes, hash them, and compare to the published output_hash. Independent implementations have reproduced the format, and a self-consistent file proves it agrees with itself. Coverage, however, requires claiming steps to be resolvable to records or reported as absent. Verification tools that only report a flat green checkmark without addressing these checks have not fulfilled the necessary verifications.
Rewritten by urgent.news; may contain errors — check the original.
Source: dev.to