The reason we mostly default to separating concerns of data and code is because of all the headaches it avoids. One item on my research wishlist is to bring more constraints like this provable to language models.