Without re-testing their entire suite of cards for every new card review (which is cost prohibitive), performance changing from updates would make the comparisons between cards less useful, as it cannot be determined if the newer card being tested is better or worse purely on the merits of the hardware itself, since newer software may be artificially making it look better or worse than the tested cards that came before, and thus the actual integrity and usefulness of the testing comes into question.

They are trying to assemble a like-for-like dataset that doesn't require their entire catalog of cards to be regularly retested to ensure that it remains like-for-like. Keeping all the software the same across tests ensures that they can add new data piecemeal and still retain an apples-to-apples comparison.