The public-release-only arrangement above differs from the latter two, which form external groups to explain public material or control later release. OpenAI's 2018-05-03 debate proposal instead has opposing agents argue to a limited judge, aiming to expose assessable disputes. The contrast concerns who constructs an explanation, who challenges it and who decides; these are different review designs. Debate's restricted demonstration used handwritten-digit images and a classifier judge, not scientists reviewing unfamiliar research. Neither that demonstration nor the 2026 group announcement establishes which design improves scientific judgment. Whether adversarial agents expose errors or merely persuade the judge remains a scientific-oversight research question.

Evidence status: analytical_design_comparison_with_restricted_demonstration.