Step-by-step benchmark of Claude Fable 5, GPT-5.6 Sol, and GLM 5.2 hallucination rates and citation patterns. Learn to score factual accuracy for answer-engine