AI training on its own output is (probably) fine if it’s validated, like a lean-verified proof to Navier-Stokes.

Because it being validated as correct resolves the main issue with incestuous training, which is compounding error.