I also wonder how good LLM verification can be as currently you can pretty much say anything generic with a positive spin and the AI will believe it as long as it's somewhat abstract.