The degree to which a model's intermediate reasoning steps logically support and justify its final answer.