Towards Verified Code Reasoning by LLMs