Idea provenance: the product counterexample

Assessed on October 10, 2026, against repository commit fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb. This note concerns the ancestry of the argument and the public discovery record. Correctness assessments remain in the separate panel commentary.

The evidence supports substantial continuity with classical convex geometry. The paper's distinctive advertised contribution is an explicit Cartesian-product witness and a quantitative extension. It acknowledges an earlier negative answer to the simplex-maximum conjecture. I found no case-specific discovery sequence in the inspected public records. Manuscript, introduction, repository disclosures.

Primary sources for the table: Brannen's publisher abstract, Ludwig's author-hosted paper, Saroglou's preprint, Feng–Hu–Liu–Xu's full text, Henk's publisher abstract. These are mathematical antecedents, not evidence that the generating model accessed each source.

The construction can be understood through a short mathematical roadmap:

flowchart LR
    A[Classical facet formula] --> B[Separate the facets of a Cartesian product]
    B --> C[Normalized volumes multiply]
    D[Known simplex values] --> E[Compare products of simplex values]
    C --> E
    E --> F[Select the stated dimension-twenty witness]
    C --> G[Repeat a simplex block]
    D --> G
    G --> H[Compare exponential growth rates]

This diagram reconstructs the dependency structure of the published argument, not the generating model's chronology. Its sources are the manuscript's Proposition 3, Proposition 5, Theorem 1 and Corollary 6. Native manuscript.

The human antecedent deserves particular attention. Feng–Hu–Liu–Xu's §4 passes from section/projection estimates to polytopes using Minkowski's existence theorem, obtains a strict comparison through a geometric perturbation, and lifts the construction to higher dimensions using pyramids. That documents an earlier counterexample claim and its distinct published construction; it does not establish that their route generated the model's product idea. Their archive entry was uploaded on August 21, 2026, before the selected manuscript's September 24 date. Those public dates do not identify when the model first discovered its construction. Archive record, full paper, §4.

The public process record has a clear limit. The selected paper belongs to family 088. The released abridged reasoning summaries cover ten other families; there is no dedicated summary for 088 in the published inventory. OpenAI describes the collection as produced predominantly by one unreleased internal model, with an average of three hours of thinking compute per result. This is a collection-level average, not a measured duration for this manuscript. Family catalogue, README, released summaries.

A keyword screen of the text extracted from all ten summaries found no selected-paper title, Brannen reference or family-088 designation. The Mahler summary contains related projection-body terminology, but that overlap supplies no discovery record for this paper. The selected source path has one public commit in the inspected ancestry, dated October 6, 2026; a public source commit is not a record of the earlier internal discovery. Detailed query scope and hashes are saved in the process notes.

My interpretation is that the written argument demonstrates useful structural recombination: exploit a product operation, reduce geometry to factor values, then make a concrete parameter choice. A specialist could recognize that strategy without adopting the earlier perturbative construction. The manuscript does not expose a large search, a long succession of failed attempts, or an unresolved classical lemma overcome by exceptional endurance. This is an interpretation of the finished argument, not a claim about the model's private process or a comparison with measured human effort.

To distinguish independent discovery, adaptation of a supplied idea, retrieval from literature, and persistence through many attempts, the missing evidence would be the original problem prompt, supplied references, chronological attempt/tool records, and the point at which the product construction first appeared. A citation in a finished paper cannot distinguish training knowledge, retrieval during discovery, material supplied by an evaluator, or a reference added during writeup. A released abridged summary could narrow some of these questions, but it would still be a selected account rather than a complete experimental record.

For the referee workflow, this case motivates an additional provenance record alongside the dependency graph: classify each idea as an explicit antecedent, a close methodological analogue, a contribution claimed by the manuscript, or unresolved; attach exact source locations and dates; and keep mathematical ancestry separate from documented acquisition by the model. The current evidence leaves the exact prior attribution of the product identity and simplex-product witness open.

Supporting artifacts: primary-literature notes, public-process notes, machine-readable provenance graph, and source hashes and evidence manifest.