Debate is proposed as scalable oversight: models argue and a human judges. In practice, does debate actually surface truth, or does it reward persuasiveness over correctness? Has anyone run it at meaningful scale?