it's a bit unclear why a model works best here. in short - smart diffing is edge case hell and you'll never capture all of them