It cannot be solved this way because it's a people problem - LLMs are like people, not like classical programs, and that's fundamental. That's what they're made to be, that's why they're useful. The problems we're discussing are variations of principal/agent problem, with LLM being the savant but extremely naive agent. There is no probable, verifiable solution here, not any more than when talking about human employees, contractors, friends.