It's just an llm looking at every message and deciding if it's good or bad.