> Because it's amusing to loop this kind of criticism through a model

Maybe it could become a general pattern, to have an agent whose task is just to deny the output validity. GANs are a very successful technique, perhaps it could work for language models too.