This is what we're doing. I think it's not only possible, but also necessary for applications beyond trial-and-error generation.

Ditto, this seems to have some parallels to the neurosymbolic ideas being explored by Lab V2 at ASU

LEAN