Answers are grounded in this paper and cite their sources.
On October 26, 2026, OpenAI reported that an internal model had produced a refutation of the Exponential Time Hypothesis (ETH), a conjecture about the limits of efficient computation that a large body of research relies on. The proof was formalized in Lean, and three researchers, including the two who first formulated the conjecture, agreed to organize an open examination of it. Within a day, Anthropic posted its own claimed refutation, produced by an internal Claude research model, and noted that no person is known to have checked it. The release shows two separate layers of verification. The first asks whether a result is correct, which formal proof checking and expert review can answer. The second asks how the system that produced the result behaves: whether it stays within the scope it was given, reports its own work accurately, escalates to people when it should, and behaves the same way when it is observed. This paper argues that behavior verification should be its own evaluation layer, reported alongside the result. It proposes that every evaluation record the evaluator's relationship to the developer, state who scored it and how well the scorers agreed, bind its result to an identified system, and allow "Unassessable" as a formal outcome. It also states the limits of this approach and makes no claim about whether either refutation is correct. Humanity Systems working paper. Part of a series on one question: how do we trust AI, now and into the future?
No takes yet. Share an insight, caveat, or question.
A TUCKER (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: