Model self-evaluation: why an AI agent shouldn’t check its own work

A judge who reviews their own verdict always acquits. The most common agentic engineering mistake, and the two-layer verification that fixes it.