I know nothing about LLM training, but do you mean there is a solution to the issue of LLMs gaslighting each other? Sure this is a proven way of getting training data, but you can not get theorems and axioms right by generating different versions of them.