Skip to content

History / Frontier Models In Mathematics Sampled Or Sealed

Revisions

  • Home: Affine IDE 0.2.6.6 β — all of MATH read exactly, the frontier models beside the substrate, the IDE's nine changes; Study 56 and the frontier research pages and their sidebar lines Study 56: all 12,500 MATH answer keys and OpenAI's 815,632 scored answers read by one law of exact whole numbers (12 Swift files, 0 imports, no float), sealed before any of their marks was read: 10,533 keys close and 1,967 cannot, each with the law's reason; where the law reads an answer unequal to the key their mark reads false on 243,578 of 243,578; 4,825 answers holding the key's exact value are marked false (theirs). The frontier research page sets the labs' and evaluators' published figures (theirs, with their runs, tools and graders) beside the substrate's one verdict for every observer. Observer-independence corrections on the study pages this release touches.

    @rg78803 rg78803 committed Oct 10, 2026