Taming the Centaur(s) with LAPITHS: a framework for a theoretically grounded interpretation of AI performances