We asked 5 AI models the same 24 questions. They agreed on the facts — and split on almost every judgment call.