That's a great question. They already do some sort of post-model filtering and output-checking, right?