AI agents can sound strategic while reasoning from events that never happened

I built three versions of an AI Werewolf game. The first agents made confident accusations from behavior that had not occurred. The useful fix was not more polished prose. It was an environment that could contradict them.