It is hard to prove that the model doesn't recognize the tests and reproduces the memoized code. It's not a clean room.