That would be possible for purely funcional languages, but the mainstream ones allow side effects that you can't quite predict. Remember training a model is jus

That would be possible for purely funcional languages, but the mainstream ones allow side effects that you can't quite predict. Remember training a model is just teaching a machine to memorise statistical occurrences. Anything the model doesn't anticipate can be catastrophical. I could see something like control engineers do, use machine learn to help tune a controller. In this case, use a model to generate an aproximation code translation with accompanying degree of precision.