Yeah, exactly this. You would really want to pursue orthogonal methods for robust unlearning, so that you can still use quantization to check that the other methods worked.