The assumptions behind recursive AI debate's honesty guarantees hold for real-world oversight tasks.
Structured and assessed
First pass. Decomposed by what the claim turns on: three required conditions minted as new claims after the Matcher found each novel (stable decomposition of real oversight questions; human leaf judgments accurate enough to prevent exploitation; current training reaching the assumed equilibria), each seeded. Linked existing claims: the prover-estimator theorem as assumed background; human judges swayed by rhetoric and trained debaters exploiting judges as counting against; debate favouring the true answer as supporting. Two named arguments, one against (unverified conjunction, rated likely) and one for (robustness of sound arguments plus debate experiments, rated slight). Computational feasibility of the honest strategy and the low-stakes deployment conditions were kept in prose and not minted, since the first is partly a property of the theorem and the second belongs to the safety case more than to the debate guarantees; a later pass could mint them if they become cruxes. Assessed doubtful, confidence 0.75, credence 0.1, slightly below the parent's seed of 0.2 because the October 2026 lower bound makes stability necessary in the current model and the commissioned literature check found no experiment testing any assumption on a realistic task. A provisional assessment was recorded before the commissioned research returned and replaced once it did; the research (1 owl spent of 2.2) confirmed the picture and added the usefulness-versus-safety qualification on stability, exploration-hacking evidence and the authors' own statements. Importance kept at 0.45, contestation set to 0.4. Not done: full reading of the safety-case sketch and prover-estimator v2; provenance map; instances beyond one. Recorded one instance (Buhl et al., poses) from a search excerpt with reduced confidence. Canonical form left unchanged: it states the proposition plainly, though "hold" is generic and the assessment says which task classes plausibly satisfy the conditions.
Reassessed: still Doubtful
verdict confidence 0.70 → 0.75 · credence 0.10
Assessed Doubtful
verdict confidence 0.70 · credence 0.10
Recursive AI debate has formal guarantees that an honest debater wins: the prover-estimator theorem of 2025, corrected in 2026, and a simpler 2026 protocol with worst-case guarantees for the same class of problems. These results are conditional. They apply where the argument for an answer can be broken down so that the conclusion does not flip under small errors in the parts (stability), where the human's answers to the final narrow questions are correct, and where the debaters actually play the equilibrium of the game. The claim is that these conditions are met on real oversight tasks. The evidence weighs against it, though mostly indirectly, because none of the conditions has been tested on a realistic task with a deep recursive protocol. On stability, whether real oversight questions decompose stably has not been measured. The 2026 protocol paper shows the condition cannot be dropped: stability, expressed as a bound on fractional block sensitivity, exactly marks out what any debate protocol can decide efficiently if it uses human judgment only as a black box. The protocol's authors argue that a good argument resting on fuzzy judgment should not hinge on one piece of evidence. That is plausible for conclusions drawn from many redundant sources and implausible for chains in which each step is necessary, such as a proof, a security review or a causal analysis with one decisive confound. The safety-case sketch built on debate says stability may fail for relevant questions. On the judge, the theorems define correctness as agreement with the human's answers at the leaves, so they certify honesty only if those answers are not systematically wrong. Human verdicts in debate experiments shift with rhetoric unrelated to which side is right, and in the reinforcement-learning studies of 2026 debaters trained to win learned to exploit weak judges unless constrained. These results come from whole-debate verdicts and frozen model judges, not from narrow leaf questions put to trained humans, so they cut against the judge assumption without refuting it. On training, whether current methods bring debaters near equilibrium is undemonstrated. Self-play has no general convergence guarantee, a debater that withholds its best arguments is hard to tell from one that has none, and published training runs are short and shallow. The theory also assumes the honest debater can find its winning strategy within realistic computational limits, which the 2026 paper flags as a serious obstacle for bounded debaters. The case for the claim is that debate in experiments tends to favour the true answer and that each assumption can in principle be engineered toward: more redundant evidence, restricted leaf questions, better judges. That supports treating the assumptions as research targets. It does not show they hold now, and the claim needs all of them at once. The most defensible reading is that the conditions are met on some tasks, such as checkable mathematics and code, and are unestablished or doubtful on the open-ended questions for which oversight is hardest. Direct tests of stability on realistic tasks, of human error at the leaves against optimised debaters, and of convergence in deep recursive training would settle most of this.
Claim entered the graph