What we want could be a complexity metric corresponding to a set of lines. I'd say that's something that could be ML-learnable with the right dataset. Given a

What we want could be a complexity metric corresponding to a set of lines. I'd say that's something that could be ML-learnable with the right dataset. Given an answer from stackoverflow, ask an average developer to identify what was the question from a set of say 5. The average time taken to provide a right answer would be the complexity score. You could use a transformer trained on code, and make it transfer-learn a function approximating that score.