How do you validate this kind of work to weed out any confabulating by the LLMs?