By requiring the LLM to prove your requirements true mathematcially.
openai has retracted many of their lean proofs for the big drop of supposed novel mathematical solns. if frontier ai screws it up , how does this inconsistency manifest in weaker models?