Completely agree. It currently tends to write unsafe, error-prone code. The next step is to figure out how to rein it in, either with new techniques or rejectio

Completely agree. It currently tends to write unsafe, error-prone code. The next step is to figure out how to rein it in, either with new techniques or rejection sampling from a large set of possible outputs.