what kind of feedback loop are you looking for here? i think Dependent Haskell is is interesting here because most models already have Haskell represented in the training set. same deal w/ Liquid Haskell: existing models are competent with the base language & now you get SMT solver feedback.
