A visible chain of thought can help with monitoring without faithfully explaining an answer. I would trust LLM reasoning more when the product preserves evidence, alternatives, and state changes that another person can audit.
markhuang.ai/news/llm-shows-work-hi…
