What happened
The Yang-Mills Group repository says it collects submissions on pure Yang–Mills theory, the group’s independent checks, and its assessments. Its first package concerns OpenAI’s submission dated 28 October 2026; the group’s assessment is dated 10 November 2026. The repository describes its conclusion as an opinion: the submission proves that a relativistic quantum field theory satisfying the Wightman axioms can be constructed as a scaling limit of four-dimensional Wilson lattice gauge theory for any simple compact gauge group.
The assessment explicitly limits that conclusion. The repository says the submission does not prove that the infinite-volume limit is unique and does not prove a mass gap. Those are consequential omissions relative to broad descriptions of a solution to the Yang–Mills existence and mass-gap problem. The repository presents the result as one submission and one assessment, with later submissions to receive separate directories and assessments.
The October package contains the original Lean sources, a combined acceptance theorem, bridges to prior relevant formalizations, and a separate operator-product-expansion check. The repository says its historical evidence records what was built and which Comparator checks were run, including their limitations. Its documentation points readers through checks covering lattice law and scale, the construction chain, operator content, Osterwalder–Schrader axioms, a translation Ward identity, a leading trace law, and OPE bounds. The project distinguishes formalized results from explanatory material and provides commented examples beside compilable statements.
For reproduction, the repository pins Lean, Mathlib, Comparator, and other dependencies. It gives an integrity command, python3 scripts/verify.py integrity, and directs readers to prerequisites and reproduction instructions for a full build and Comparator replay with python3 scripts/verify.py all. It notes that the recorded macOS verification mode does not provide Linux/Landrun isolation. The repository lists Apache-2.0 for its code and documentation, subject to preserved upstream notices. It says posting a submission does not endorse or certify its claims, and the group does not promise to review every submission.
What the archive adds
New Runtime’s earlier coverage on 2026-09-09 described OpenAI’s claim to have solved the Navier–Stokes Millennium Prize problem. This is a new event in the same broader thread: an assessment of a separate mathematical submission, with a stated boundary around what the proof does and does not establish. The earlier note is not a verification of this Yang–Mills work.
Trend signal
This episode supports the assumption that expert review of AI-associated mathematical proofs will eventually examine more than one headline result. The Yang-Mills Group presents a dated assessment, checks, and reproduction materials while preserving the distinction between the submitter’s source and the group’s additions. See New Runtime’s assumption on expert review of AI proofs.
Connections to other events
New Runtime’s earlier coverage on 2026-10-22 described Codex and Claude performing an ALEPH measurement under physicist direction, from raw data to a paper, and treated that as a workflow claim rather than a precision-physics claim. This assessment concerns a different kind of scientific work: a mathematical construction accompanied by formalized checks. Both cases make the distinction between a system’s contribution and the human or independent process that evaluates the resulting claim important.
Why it matters
For teams building with AI, this is a useful model of how to package a consequential claim for scrutiny: preserve the original submission, state the proposition narrowly, separate formalized results from informal arguments, pin dependencies, and give readers a route to rerun checks. Those practices make it easier to identify exactly what a tool established and where human mathematical judgment remains necessary.
The repository also illustrates why “verified” is not a single status. A build can be reproducible while the scope of the theorem remains narrower than a popular summary suggests; a committee can assess a proof without claiming to understand every step; and a formalization can have explicit assumptions or gaps. Product teams can borrow this discipline for AI-generated code, analyses, and scientific outputs: define acceptance criteria in advance and publish evidence tied to those criteria, rather than treating a successful model response as validation.
Limits and uncertainty
The group frames its conclusion as its opinion, not as certification. The repository does not establish the full Yang–Mills problem: it explicitly says uniqueness of the infinite-volume limit and a mass gap remain unproved by this submission. Its own description says the historical checks have limitations, and the recorded macOS verification mode lacks Linux/Landrun isolation. The repository’s public description does not establish that every independent mathematician has reviewed or reproduced the work, nor that the result has been accepted as a solution to the Millennium Prize problem.
Skill to build now
For an AI-assisted proof or analysis, write a one-paragraph claim boundary before asking anyone to review it: list what is proved, the assumptions and formalized components, and the tempting but unproved conclusions. Then package the original artifacts separately from your checks, pin the toolchain, and include a clean reproduction command. This week, apply that format to one model-generated technical result in your own workflow.
Reader takeaway
The meaningful update is not that an AI has conclusively solved the Yang–Mills problem. It is that an independent group says a particular construction claim is supported while explicitly excluding uniqueness and the mass gap. Read the theorem’s boundary and the verification record before repeating the headline.
Open question
Will further independent mathematicians reproduce and assess the package, and will later submissions or checks address the two results the group says this submission does not establish?