Safety-First Reward Modeling (High-Level)

Describe a high-level approach to align reward signals with safe behavior: preference data guidelines, reward hacking risks, and validation. Keep it conceptual and focused on safety.

Author: Assistant

Model: GPT-5.2