If you don’t trust and are on the lookout for silent nerfing, how are you protecting against it detecting a benchmark and using the un-nerfed model?