How would you design a system to automatically evaluate a new LLM checkpoint for harmful behavior before deployment at Character.AI?
