A repository can make a mathematical claim inspectable without making it accepted knowledge. OpenAI's October 6 release highlights that gap: the retained account lists 719 manuscripts grouped into 372 families from an unreleased internal model. Supporting arguments and alternative proofs can share a family, so the manuscript count is not a count of independent breakthroughs.
The practical response is a result-level record. For each important claim, connect the exact theorem, assumptions, provenance, formal artifact and responsible human review. This makes useful results easier to identify without asking the mathematical community to treat publication volume as evidence of correctness or significance.
Keep the search denominator visible
OpenAI reports roughly 4,000 attempted problems before aggregation and significance filtering. That context matters, but it does not supply the difficulty distribution, every failed approach or a reproducible selection rule. Do not divide manuscripts by problems and call the result a general success rate: families, variants and filtering have different units.
The company released 10 abridged reasoning summaries, not a complete production history for every manuscript. Aggregate compute figures likewise do not replace result-level prompts, model identity, elapsed work and revisions. An inaccessible model prevents outsiders from rerunning the same system or testing neighboring questions under matched conditions.
Verify what the formal code actually says
The repository reports formalization for approximately 42 percent of top-line results. That fraction refers to top-line claims, not all 719 manuscripts. Lean can check that encoded steps follow from declared definitions and assumptions; it does not automatically establish that the encoded statement is the theorem advertised in prose.
A separate analysis of the earlier Navier–Stokes work identified translation discrepancies without declaring the entire natural-language argument false. One bound used four additional derivatives in prose but five in the associated formal result. The difference illustrates why compilation and semantic fidelity are separate checks, not interchangeable badges.
Map every important claim to reviewable support
Ask which Lean declaration supports each theorem and intermediate statement, what assumptions changed, and which parts remain unformalized. Preserve machine-readable links between prose and code, build instructions, dependencies and known limitations. A reviewer should be able to distinguish a corrected exposition from a changed mathematical claim.
The independent AGMAI group recommended responsible release practices; consultation was not certification of the batch. Its advice distinguishes work understood by a human author from results still awaiting understanding. Peer review, seminars, questions and accountable correction remain relevant even when a formal checker accepts part of the argument.
Fund understanding, not only generation
Correctness, originality, usefulness and explanation are different outcomes. A true theorem may duplicate prior work or contribute little; a valuable method may require substantial exposition before others can use it. More manuscripts can increase review burden faster than they increase shared knowledge, particularly in narrow fields with few qualified experts.
OpenAI says it will support workshops, conferences and programs. Their independence, structure and scale remain unspecified. Useful support would produce expository work and researchers able to defend the mathematics, not merely more publicity. Independent allocation of community support also helps preserve researchers' freedom to choose questions outside a lab's evaluation agenda.
Watch individual claims survive review, formal mappings improve and other mathematicians reuse the methods. Retain visible revision and withdrawal histories rather than silently replacing failed claims. The strongest evidence of progress is a growing body of independently understood results; neither a large repository nor a partial formalization percentage alone establishes that achievement.