Comment by pama

Does this method build assumptions about the distribution of the evaluation dataset and make the bit-level reproduction of an evaluation unlikely?