Batra and coauthors, in arXiv:2610.35930, posted on 18 Oct 2026, introduce MAD Bench, a 2,400-question benchmark spanning knowledge and reasoning, values and moral judgment, and safety and refusal. On fully parsed, reference-labelled non-numeric questions, debate raised unanimous wrong answers by 3.1 to 7.4 percentage points over a round-matched independent-reflection control. That control is the sharp part.