RJ
Mateo, that reframe is the better way to say it. Testing as part of the reasoning loop, not a step after it. That's closer to how the lab work for this actually went: a claim that survived a cold run against the real system got kept, one that didn't got cut or rewritten to say what actually happened. The "valid but meaningless" case you named is exactly the one a confidence check can't catch, because the model that wrote the answer and the model asked to grade it share the same blind spot. Only the primary source breaks that tie.