>but it doesn’t have to be correct, accurate, or ...

why we don't do GAN here, ie. second model verifying correctness/accuracy/etc. ?