An audit trail for AI decisions records, for every automated or AI-assisted action: the signals read, the reasoning followed, the policy applied, the human who authorised it, and the outcome — in a tamper-evident, hash-linked ledger captured at decision time, not reconstructed afterwards. MAIA seals this proof packet before consequential actions execute, which makes the audit trail a precondition of autonomy rather than a hope.
Logs are not an audit trail
Ordinary logging serves the operator on a good day; an audit trail serves the institution on its worst day, under adversarial review. The differences are structural: typed records instead of free text, hash-chaining so alteration is detectable, signatures so actions are attributable, and replayability so a determination can be re-run exactly as it ran.
The capture moment matters most. Evidence assembled after an incident is testimony; evidence sealed at decision time is a record.
What auditors actually ask
Three questions recur in every serious review: who authorised this class of action, what did the system know when it acted, and can you show me the same answer twice? A decision ledger with lineage answers all three from one artefact — which is why evidence-first architecture tends to shorten audits rather than complicate them.