Can this approach be made more robust so it could be used in critical programs? Is there a way to formally verify performance of an approach like this?