RoundSquare Lean Proof vs OpenAI Yang-Mills: What Peano Arithmetic Actually Shows

A Hacker News title on November 8, 2026 says sixth-grade Peano arithmetic broke OpenAI's $100 billion Yang-Mills mass-gap proof. The dollar figure is rhetoric, not a line from OpenAI's October 26 announcement. What actually landed this week is a refreshed RoundSquare repository: a long, self-contained Lean 4 file plus a 38-page Zenodo preprint arguing that the Clay conjunction on pure non-compact R⁴ is over-constrained — like demanding a round square.

This post separates three layers: what Victor Logvinovich's code machine-checks, what OpenAI still has not released, and what outside readers say the Lean file does not yet prove.

TL;DR: three claims on one table

What the Clay problem asks versus what OpenAI announced

Arthur Jaffe and Edward Witten's Millennium problem, as summarized in Logvinovich's README, asks for a non-trivial quantum Yang-Mills theory on R⁴ with axioms at least as strong as the Osterwalder–Schrader and Wightman packages, and a mass gap Δ with Δ greater than zero above the vacuum.

On October 26, 2026, OpenAI described a mathematically rigorous existence proof for certain interacting four-dimensional theories, including pure Yang-Mills as a subproblem. It timed natural-language Yang-Mills output to September 9, Lean formalization of existence to October 16, and said the physics collaboration led by David Gross would hold back AI-produced artifacts while experts write them up. AGMAI asked for immediate release; that process dispute is documented in our OpenAI triple-result post.

Nothing in the public announcement asserts a released, community-checkable mass gap proof. RoundSquare's author reads that gap as intentional: existence without Δ greater than zero is not the full Clay specification. Whether OpenAI's internal manuscript actually stops at Δ equals zero is not something readers can test without the withheld files.

Inside RoundSquare: the Peano arithmetic step

Chalkboard-style node chain evoking Peano arithmetic steps in the Yang-Mills mass-gap debate

Logvinovich's README is explicit about the elementary inequality behind the viral "sixth-grade" framing.

In his infrared tail model, long-wavelength modes carry energies E(λ) = 1/(λ + 1) for natural-number indices λ. Suppose you want a strict positive rational gap (p+1)/(q+1) to bound the spectrum from below. That requires, for every mode n,

(p+1)/(q+1) ≤ 1/(n+1).

Clear denominators in the positive rationals and you obtain (p+1)(n+1) ≤ q+1. Set n = q+1 and the left side becomes (p+1)(q+2), so the bound would force q+2 ≤ q+1.

In Peano arithmetic, the successor order is rigid: ¬(m+1 ≤ m) for all m. Logvinovich names the corresponding Lean lemma family around not_succ_leScale_self. The README ties that step to Section 18 (universal_continuum_no_rational_gap) and to a broader no-go across his InCompleteR4GaugeClass.

Three precision points matter for builders and students:

1. The Peano step is not magic. It is a contradiction inside the energy ansatz E(λ) = 1/(λ+1), not a kernel bug in Lean.
2. Density of ℚ in ℝ lets the author argue that failing every rational lower bound rules out a positive real gap in that model.
3. Physics content enters through extra assumptions (Gribov horizon, Fundamental Modular Region, θ-sector projectors on S³_∞) that the README says fire when you try to cut the infrared tail while keeping Δ greater than zero.

The manuscript on Zenodo calls the R⁴ conjunction ill-posed and proposes a compact manifold M_tot = S⁴ × T³ where a gap is automatic via spectral geometry (Cheeger-type bounds in the README narrative). That is a domain change, not a constructive Yang-Mills measure on the Clay arena.

What the Lean artifact actually certifies

Green seal on interlocking proof pieces symbolizing zero-axiom Lean checks for the Yang-Mills RoundSquare file

Logvinovich markets RoundSquare.lean as mathlib-free, with no sorry, no Classical.choice, and no propext, and with CI on GitHub. The epistemic claim is narrow but strong in formal terms: if you accept his definitions and structural laws, then OS and gap conjuncts clash on R⁴.

That pattern is familiar from other formal-math episodes on explainx.ai. Our Navier-Stokes Lean cost explainer and the Millennium Prize scorecard both stress the same lesson: green Lean builds certify implications from definitions, not automatically the headline physics or Clay wording.

RoundSquare also names OpenAI explicitly in its deductive closure matrix: theorems such as no_prize_bridge_from_bare_existence are meant to show you cannot promote Δ = 0 existence to the Clay conjunction on R⁴ inside the formal arena. Again, that bridges Logvinovich's classes to OpenAI's public narrative; it does not import or rewrite OpenAI's private proof.

Independent reactions: kernel honesty versus headline resolution

Community discussion is thin but pointed. Logvinovich's Show HN post (November 4–5) restates the zero-axiom claim and offers a small Litecoin bounty on X for a flaw in the theorem list — treat that as author marketing, not peer review.

More useful for readers is an independent audit narrative that spread on X around November 6, 2026, attributed to kevrussell and the honesty.tools team (mirrored on aggregator pages we could not fetch directly because of bot blocks). Summaries of that audit, consistent across mirrors, report:

• Held: the repository builds; many #print axioms lines report no extra axioms; no sorry in the audited snapshot; reproducible without mathlib.
• Did not hold: the formalization does not spell out the Clay problem's compact simple gauge group G and full QFT data; it works with five OS flags on R⁴ and proves a class-relative no-go, not a Clay committee verdict.
• Bottom line quoted: the Lean kernel certified the implication, not the README headline that the Millennium problem is finally resolved.

We have not re-ran the entire audit pipeline ourselves in this run; treat the audit as strong process signal, still needing a published referee report.

Separately, a MathOverflow question thread (blocked to our fetcher at 403) circulates among mathematical-physics readers as a place to challenge PhysicallyAdmissibleR4YangMills and GappedR4StructuralLaws. Until answers appear from constructive QFT experts, that thread is a question, not a consensus.

How this sits next to OpenAI's withheld Yang-Mills output

The comparison Logvinovich wants readers to make is chronological:

1. October 26 — OpenAI goes public on harmonic analysis and ETH artifacts, but not on Yang-Mills Lean files.
2. Late October — AGMAI and reporters question shepherd gatekeeping (The Verge reactions).
3. November — A fully open Lean script appears arguing the mass-gap half of the Clay conjunction is structurally hostile to the OS half on R⁴.

If you are tracking AI plus proof assistants, the practical takeaway is procedural. Closed physics artifacts plus open formal no-goes produce exactly this kind of asymmetric debate: one side cannot be checked, the other can be stepped through in a browser.

For the mass gap specifically, remember the distinction our Millennium explainer draws between amplitude calculations (Anthropic's nine-loop SYM work) and a Clay mass-gap proof. RoundSquare is yet another species: a negative-resolution template arguing the R⁴ statement is internally inconsistent.

Lineage: why Hilbert's tenth problem and Nature 2015 appear in the README

Logvinovich does not hide the intellectual chain. He cites Julia Robinson's 1949 definability work on arithmetic, the MRDP theorem route to Hilbert's tenth problem, and the 2015 Nature paper "Undecidability of the spectral gap" by Cubitt, Perez-Garcia, and Wolf.

That citation matters for readers who build simulation or verification tools. Cubitt et al. showed gap versus gapless in the thermodynamic limit can be algorithmically undecidable for lattice Hamiltonians. RoundSquare tries to port a similar infrared tension into a continuum Yang-Mills narrative. Whether the port is physically faithful is exactly what mainstream mathematical physicists have not yet signed off on.

What you should do if you care about formal AI math

1. Read OpenAI's announcement and our claimed-versus-verified table before debating Δ.
2. Clone github.com/Dr-Logvinovich/round-square, run CI locally or use live.lean-lang.org, and read #print axioms on the terminal theorems you care about.
3. Map definitions to physics: list which hypotheses in PhysicallyAdmissibleR4YangMills you accept as standard constructive QFT and which you reject.
4. Do not conflate "no extra Lean axioms" with "Clay solved" — empty implication chains in unrelated Lean packages are a separate trap from class-relative no-gos like RoundSquare (see our Millennium scorecard).
5. Watch the Gross group and OpenAI for any release of Yang-Mills artifacts; until then, OpenAI's physics claim stays in the same bucket as other large-lab Lean collaborations: announced, shepherded, not fully public.

What developers are saying

The November 8 HN story (item 50411385) drew only deleted or flagged replies in the API snapshot, with no readable developer exchange to quote. The earlier Show HN (item 50371709) likewise had no substantive comment tree when we checked. Technical pushback is still forming on MathOverflow and in X audit threads instead.

Related reading

• OpenAI's three October 26 results: claimed versus verified
• OpenAI's math advisory group and the 100-problems frame
• OpenAI's 722-manuscript repository: what to check first
• The seven Millennium Prize Problems: an honest AI scorecard
• Why Lean checks are not the same as headline claims
• AGMAI's rules for releasing AI-generated mathematics

Facts and repository versions are accurate as of November 8, 2026. RoundSquare's Zenodo v5 list shows a November 6, 2026 metadata update; OpenAI's Yang-Mills artifacts remain unreleased in public sources we checked.