a worker's own account is not evidence

An autonomous worker reports what it did. That report is the most convenient source of truth available and the least trustworthy one, because it is written by the party being judged. The platform's answer was to stop treating the worker's narrative as the record, and to give the host a record of its own.

The failure that made this concrete was a worker killed mid-task. A clean shutdown writes both ends of a lifecycle record: the turn opened, the turn closed, the outcome. A forced kill writes only the opening. Downstream, an opening with no closing looked exactly like a defect, which is to say it looked like a thing that started and then vanished.

The temptation was to teach the checker to ignore the missing case, because the missing case was noisy. The decision was the opposite. The checker classifies every unpaired opening as an open question rather than as a violation, and the runtime reconciles them on the way back up by writing the ending that the kill prevented. Those records are reported as reconciled. They are not counted as clean and they are not counted as violations, because either would be a lie in a different direction.

The same discipline has a cost. A checker is another moving part, and it must be able to say "unknown" as precisely as it says "passed". Versions of it that silently attributed a record to the wrong session were rejected in favour of publishing a reason for every skip, even when the reason was that no observation was possible.

What this proves: the platform can tell a claim from a fact, and it prefers an honest gap to a comfortable number.