Most "audit logs" are just a table someone can UPDATE. That's a record, not evidence. If the history of who approved what can be quietly rewritten, it proves nothing.
The direct answer
An audit trail is tamper-evident when every entry is hash-chained to the one before it, written atomically with the state change it records, and stored append-only with no update or delete path. Any change to a past entry breaks the chain and is detectable.
Hash-chaining, briefly
Each event includes a hash of the previous event. Alter an old entry and every subsequent hash no longer matches — verification catches it immediately. You don't have to trust the storage; you can check it.
Atomic by co-commit
A decision and its audit record must land in the same transaction. If they don't, you get drift: a status with no explanation, or an audit entry for a change that rolled back. Seal co-commits them, so that gap can't exist.
Append-only, enforced twice
The log has no update or delete path in application code, and a database guard rejects mutation attempts outright. Append-only isn't a convention here — it's a property of the system.
Why it matters for agents
When an autonomous agent takes a sensitive action, "who approved this and when" is exactly the question you'll be asked later. A tamper-evident chain answers it with proof, not with a table you hope no one touched.