The trick is, you need two LLMs, one which always lies, and one which always tells the truth. Then you ask either LLM whether the other LLM would say it's reliable.