The core question is whether a code-writing AI's output can be trusted, and the only evidence that matters is whether that output actually works. A test suite written by the same AI is just the model agreeing with itself; it doesn't validate behavior against independent requirements. Testing by the output's own test suite is circular reasoning. What we need is an external oracle, a spec someone else wrote, to catch the systematic blind spots these systems have. Self-testing flatters the model but doesn't verify correctness. That's why the output's test suite, run against real expectations, is the honest benchmark.