Difficulty is computed from each question’s structure and recalibrated against real solve data — never asserted.