If chain-of-thought is core to reasoning, why would OpenAI not expose those tokens? Why only give a summary? What are we not supposed to see?