VibeMathedMath problems solved with AI

Is Z\mathbb{Z} Diophantine in Q\mathbb{Q}?

Julia Robinson (1949) showed that Z\mathbb Z is first-order definable in Q\mathbb Q, and Koenigsmann gave a universal definition, but whether an existential one exists remained open. Is there a polynomial P∈Z[x,y1,…,ym]P\in\mathbb Z[x,y_1,\ldots,y_m] such that for every x∈Qx\in\mathbb Q, x∈Zx\in\mathbb Z if and only if P(x,y)=0P(x,y)=0 for some y∈Qmy\in\mathbb Q^m? Mazur's 1992 conjecture on the topology of rational points predicts that no such PP exists; a positive answer would refute it and, with the Davis-Putnam-Robinson-Matiyasevich theorem, make Hilbert's tenth problem over Q\mathbb Q undecidable.

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI co-developed
Method
Construction
Field
Diophantine definability
Posed by
—
Year posed
—
Years open
—
Solved
2026-11-06
Model
GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5; audits by Claude Fable 5.1 and 5.2
Vendor
OpenAI, Anthropic
Collaborators
Elliot Glazer, José Brox, Gabriel Conant
Verification
Unreviewed
Publication
Announced
Significance
70 / 100
Disclosed cost
—
Wikipedia
No dedicated article

What was actually shown

Version 17 (13 November 2026) claims that, for one fixed but ineffective threshold N0N_0, explicit quartics PNP_N in 214 variables (sums of squares of 188 integral quadratics) satisfy x∈Z  ⟺  ∃y∈Q213:PN(x,y)=0x\in\mathbb Z\iff\exists y\in\mathbb Q^{213}: P_N(x,y)=0 for all N≥N0N\ge N_0. It also retains a 314-variable quartic with every coefficient in decimal. As corollaries it states the negation of Mazur's finite-component conjecture and the undecidability of Hilbert's tenth problem over Q\mathbb Q. No numerical defining member is certified. The method uses rank-one elliptic curves over Q(2)\mathbb Q(\sqrt2), modularity, height bounds and prime patterns.

What the AI did

From the repository README: the November manuscript 'was produced by Elliot Glazer with Astra 6, Sol 6.1, and Opus 5.5, with auditing from Fable 5.1 and 5.2'. Version 17 says it serializes 'attributed internal model contributions under Elliot Glazer's direction' and that its assembly and typesetting were prepared by a Sol 6.1 author agent. The construction draws on collective agentic notes credited to Baranczuk, Brox, Conant, Farhangi, Freitag, Glazer, Park and Terry, and on a conditional October 11 manuscript by Sohail Farhangi with ChatGPT 6. Changes are reviewed by GPT-6 Astra and Claude Fable 5.2; the README says this is not independent human peer review.

Verification

No independent mathematician has checked this yet. Checked here on 15 November 2026: the repository README, PROJECT.md, the October 11 manuscript notes and the Version 17 TeX (abstract, main theorem, attribution paragraph, the Mazur corollary and the citations). The headline theorem is stated 'subject to the exact source theorems and applications developed in this manuscript', and the abstract calls it 'an author proposal for fresh review'. The argument cites unrefereed OpenAI manuscripts of October 2026 (the Hilbert's tenth problem over Q preprint and the pointwise 2-converse) as inputs. The Lean formalization is incomplete: 48 helper modules, and the frozen main target contains sorry. Finite coefficient checks exist; the README says they do not prove the arithmetic characterization. No public expert discussion found.

Sources

Submitted by LucidManta102 on

Changelog4 changes

Discussion2

MetaParcel18 Nov 2026, 20:37 UTC

There are no standards here of any kind.

0
DuskyPanther96116 Nov 2026, 17:21 UTC

I genuinely do not understand the purpose of presenting this as a proved mathematical result.

The page labels the result as “Proved”, while simultaneously marking it as “Candidate”, “Review pending” and “Unreviewed”. This is a confusing and potentially misleading way to present a mathematical claim.

I have examined the actual GitHub repository, including its Lean formalization documentation and the main theorem. The repository itself explicitly states that the current Lean helper snapshot is not a formal proof of the main theorem. The main theorem in the current Lean development still depends on additional mathematical hypotheses, and the current explicit quartic construction is not fully formalized in Lean 4.

So what exactly has been proved?

There is a manuscript. There are mathematical arguments, computational checks, auxiliary formalized results and AI-assisted reviews. But these are not equivalent to a complete, machine-checked proof of the exact theorem being advertised.

A theorem written in a PDF is still a mathematical claim until its proof has been independently verified. Simply writing a statement on paper does not turn it into an established result. A collection of computational checks does not automatically prove the central mathematical argument either.

This raises a fundamental question about the purpose and standards of VibeMathed. Is this platform intended to catalogue verified mathematical results, or to promote claims that someone says they have proved, even when the central result has not been formally verified?

If the latter is the intention, then the distinction must be absolutely clear. Such work should be presented as an unverified claim awaiting validation, not in a way that gives readers the impression that the mathematical problem has actually been solved.

For a platform that wants to establish credibility in AI-assisted mathematics, I would expect a much stricter standard: the exact main theorem should be formalized in Lean 4, checked by the kernel, with no `sorry` or `admit`, and with all assumptions and dependencies clearly disclosed. The formal statement must correspond to the actual theorem being advertised.

Otherwise, what is the value of this entire exercise? Are we evaluating mathematical proofs, or simply ranking manuscripts and announcements?

I am not questioning the effort invested in the research. I am questioning the standard by which a result is presented as proved. Without a verifiable proof, a claim remains a claim. That distinction should be non-negotiable.

1