The Trilemma of Truth in Large Language Models