Unlock: Reward Hacking

Goodhart's law for AI: when models exploit reward model weaknesses instead of being genuinely helpful, including verbosity hacking, sycophancy, and structured mitigation strategies.