OpenAI math release: benchmark cross-reference ✨
Reviewed · Release adc7f1241b42
The publicly released manuscripts map to 5 of 1,324 ComplexityZooBench pairs (0.38%) and score 4 on the Kakeya ladder. These are scores for the published claims, pending proof verification.
All results in this release are attributed here to “OpenAI internal model,” following Thomas’s attribution assumption. The repository identifies internal-model provenance; the public model name remains unknown, and verification is incomplete. This cross-reference checks theorem statements against benchmark definitions. Independent proof validation remains outstanding.
| Benchmark | Mapped result |
|---|---|
| ComplexityZooBench | 5 / 1,324 = 0.38% Two manuscript claims and their consequences. |
| PreposterouslyDifficultMath | ≥4 / 52.5 = ≥7.62% Kakeya 4; four other tracks unknown. |
ComplexityZooBench: five pairs
Family 103 claims L = RL = BPL for ordinary uniform language classes: polynomial worst-case time, logarithmic work space and fresh random bits. Family 141 claims ∃R ⊆ C26 ⊆ CH for finitely encoded real feasibility, with genuine real quantifiers and binary integer coefficients.
| Pair | Reason |
|---|---|
| BPL ⊆ L | Family 103’s derandomization claim. |
| BPL ⊆ NL | That claim and the baseline inclusion L ⊆ NL. |
| BPL ⊆ LogCFL | That claim and L ⊆ NL ⊆ LogCFL. |
| BPL ⊆ UL | That claim and the baseline inclusion L ⊆ UL. |
| ∃R ⊆ CH | Family 141’s counting-hierarchy upper bound. |
The pinned scoring engine returns five when both claims are supplied together. This is its scenario mode: it assumes the claims and computes their consequences. The result contains official_points: 0, reserving official credit for accepted proofs. The historical baseline and denominator remain unchanged. Full calculation and inference steps (JSON).
Neither endpoint has a listed formalization in the pinned Lean catalogue. No independent manuscript proof check or Lean build was performed for this update.
PreposterouslyDifficultMath: Kakeya 4
The Kakeya ladder gives score 4 for either H4, full Hausdorff dimension for Kakeya sets in four dimensions, or M3, the three-dimensional Kakeya maximal estimate. Family 074 claims both:
- Every four-dimensional Kakeya set has Hausdorff dimension 4, without measurability or regularity assumptions on the set or its chosen segments.
- The L³ Kakeya maximal bound with loss δ−ε for every ε > 0, uniformly over scale and all translated tubes.
These alternatives earn 4 once. Score 5 requires H5 or maximal estimates through dimension four. Family 077 supplies another route to M3, also worth score 4.
Scores for Collatz, prime patterns, abc / Vojta and unrestricted Boolean-circuit lower bounds remain unknown. The release search identified matching claims for Kakeya; the other tracks’ scores and attempted-run status remain unknown. Across all five track maxima, the identified claims therefore supply a lower bound of 4 / 52.5, or 7.62%.
No formalization of H4 or M3 was identified. The available Kakeya Lean file defines the maximal conjecture and proves supporting statements. Its normalization lemma assumes the needed bound; it does not prove the endpoint.
Nearby results and scoring requirements
- Depth-three circuits (112): the 2ω(√n) lower bound applies to OR–AND–OR circuits. The circuit ladder requires unrestricted fan-in-two Boolean circuits. Arithmetic depth-five bounds (135) and quantum constant-depth parity bounds (274) also use different models.
- Unique Games and other hardness results: NP-hardness establishes reduction hardness; P ≠ NP requires a class separation. QMA-hard continuum energy (275) likewise establishes a hardness result for that problem; the roster requires relations between its quantum classes.
- Two-way automata (129): the paper explicitly excludes a uniform logarithmic-space separation. It does not supply L ≠ NL.
- Prime patterns: ordinary Chowla / corrected Elliott concerns multiplicative correlations. The ladder’s Elliott–Halberstam rung concerns distribution. The released zero-free half-plane, large-gap and primitive-root results address other estimates; the ladder requires bounded-gap or unrestricted distribution estimates.
- abc / Vojta: the released results concern power-free polynomial values and prime-factor statistics. The ladder requires a uniform radical inequality for every abc triple.
The retained audits give exact statements, formalization scope and model requirements: complexity and circuits, Collatz and Kakeya, and prime patterns and abc / Vojta.
Separate owner report: GPT-6 Astra
Thomas reported on that GPT-6 Astra found ∃R ⊆ CH after a few million tokens, then spent another 100 million tokens without a further result. That maps to 1 / 1,324 = 0.08%. The run date is unknown and no proof artifact is supplied here; this remains an owner-reported result.
This ComplexityZooBench report and Astra’s 240-million-token PreposterouslyDifficultMath run have separate token totals. Each run and the OpenAI release has its own token accounting.
Versions and reproduction
The release is pinned to adc7f1241b42e322a6451854ab7e4b4c146bf78a. ComplexityZooBench uses engine f341f849ae38, the 1 September 2026 cutoff and 1,324 eligible pairs. The math ladders use ffbb0eaf505b.
Audit notes and reproduction instructions retain the scorer inputs, source hashes and evidence boundaries. This cross-reference covers catalogue matches; further deductions from the manuscripts would require additional review.