Policy Frameworks for Transparent Chain-of-Thought Reasoning in Large Language Models