Ai Engineering 5 min read

GPT-6 Astra helped build a candidate lattice definition of the Standard Model

UCSB physicist Nathaniel Craig says OpenAI's GPT-6 Astra took the lead on the proof in his 123-page paper on lattice chiral fermions, published with a Lean 4 formalization. Here is what was proved, what Astra did, and what remains open.

On November 15, 2026, OpenAI launched a new science blog with a guest post by physicist Nathaniel Craig, A step toward defining the Standard Model on a lattice, in which he credits GPT-6 Astra with taking the lead on the detailed proof behind his paper. The paper itself, “A gauge-invariant measure for lattice chiral fermions”, went up on arXiv on October 6 with Craig as the only listed author and affiliations at the Kavli Institute for Theoretical Physics, UC Santa Barbara, and OpenAI, where he is a visiting scientist.

The claim matters because of what it targets. The Standard Model is the most tested theory in physics, yet it has no complete nonperturbative mathematical definition: nobody has written down a lattice version of it that keeps its chiral gauge symmetry exact. Craig’s paper proposes one, as a candidate, and ships a Lean 4 formalization alongside it. If the construction holds up under expert scrutiny, it would be one of the clearest cases yet of a frontier model doing the central technical work on a recognized open problem in physics.

The problem the paper attacks

Lattice field theory puts space and time on a grid so a theory can be defined exactly and simulated on computers. That works well for quantum chromodynamics, but the Standard Model’s weak force acts differently on left-handed and right-handed fermions, and chiral fermions resist being put on a lattice. The Nielsen-Ninomiya theorem, which the paper’s introduction walks through, says a lattice Dirac operator cannot simultaneously have the right continuum limit, avoid extra fermion species, and keep exact chiral symmetry.

The accepted workaround is the overlap operator of Narayanan and Neuberger, which satisfies the Ginsparg-Wilson relation and, as Martin Lüscher showed, keeps an exact form of chiral symmetry at finite lattice spacing. That leaves one hard piece: the fermion integration measure. Its phase has to be chosen smoothly and in a gauge-invariant way across all gauge field configurations, and that is only possible when gauge anomalies cancel. Lüscher solved this nonperturbatively for abelian theories around the turn of the century. For nonabelian theories like the Standard Model, the paper says the problem “remains unsolved” and that Lüscher’s abelian construction does not carry over directly.

What the paper proposes

Craig’s construction compares the overlap chiral projectors of the same theory at successively finer lattice spacings, and sums those comparisons into a measure current. Anomaly cancellation is what makes the sum converge; for one Standard Model generation embedded in SU(5), an exact identity in the fermion representation supplies the needed decay. The resulting current is exponentially local and satisfies the conditions Lüscher’s reconstruction theorem needs to produce a smooth, gauge-invariant measure.

The paper applies this to the Standard Model gauge group SU(3)×SU(2)×U(1) in all four of its global forms and for any number of generations, plus two further families of SU(N) chiral gauge theories. All of it is restricted to “admissible” gauge fields, a smoothness condition standard in this line of work. The paper runs to 60 pages plus 63 pages of appendices, and its arXiv record links a mathematical companion and Lean formalization and a two-dimensional supplement on Zenodo, with a Lean verification report among the ancillary files.

What Astra did, and what Craig did

By Craig’s account, as reported by AI Weekly, he already “had a strategy for tackling this problem, but had not been able to prove that it worked in all the cases it needed to cover.” GPT-6 Astra helped turn that strategy into a construction and “took the lead in developing the detailed proof.” The Defiant summarizes it the same way: Craig credits the model with key construction and proof steps.

That division of labor is the interesting engineering detail. The idea and the problem choice came from a domain expert; the model did the long, case-heavy proof work that usually eats months of a theorist’s time; and the Lean formalization gives a machine-checkable record of at least part of the argument rather than leaving trust to the model’s prose. It follows the same pattern as Astra’s earlier results, such as the 10 open math problems it resolved with Lean 4 proofs in August, now applied to physics where the definitions are harder to state formally.

One gap is worth flagging. The arXiv listing names only Craig and does not mention AI in its abstract, so the account of Astra’s role rests on Craig’s own description in the OpenAI post. The Lean project’s exact coverage of the main proof has not yet been independently described.

Caveats and what to watch

Craig is careful about scope. He calls the result a candidate definition and, per AI Weekly, “far more an answer in principle than an answer in practice.” Several things remain open: showing reflection positivity (needed for a sensible quantum theory), connecting the lattice definition to a continuum one, and making the construction usable in realistic lattice simulations. Practical computation with chiral gauge theories is not unlocked by this paper.

Craig has named three lattice theorists who reviewed the work privately: Maarten Golterman, Andreas Kronfeld and Yigal Shamir. No independent published assessment has appeared yet. The things to watch are a referee process or public responses from the lattice community, scrutiny of the Lean formalization and how much of the main theorem it covers, and whether OpenAI’s new science blog becomes a steady channel for model-assisted results that come with machine-checked proofs rather than only claims.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

gr.Workflow Turns AI Pipelines Into Deployable APIs

Learn how to model, debug, expose, and deploy multi-step AI pipelines with Gradio Workflow and daggr.

Ai Engineering

Astra Model Solves 10 Open Math Problems With Lean 4 Proofs

OpenAI has revealed Astra, a new AI model family designed for continuous workloads that successfully solved ten open problems in advanced mathematics.

Ai Agents

Google's /teamwork results: 1.8 million lines of Lean for the Weil conjectures, a 2013 conjecture proved, and a Rust Zircon

On November 13, Google Research published what its Antigravity /teamwork multi-agent harness produced over several weeks: a proof of the Courtade-Kumar conjecture, a machine-checked formalization of the Weil conjectures, and 800,000 lines of Rust for the Zircon kernel. Here is how the harness works and what the post leaves out.

Ai Engineering

PSI says a non-public Anthropic model proved the Aaronson-Ambainis conjecture, with a full Lean check

Physical Superintelligence says its research system, running an unreleased Anthropic model, proved a central open problem in quantum query complexity by day three of a run. The 9,454-line Lean formalization uses no added axioms; independent review has only just started.

Ai Engineering

OpenAI Claims Its Model Falsified the Exponential Time Hypothesis. The Verifying Starts Now.

OpenAI announced October 26 that an internal reasoning model produced a Lean-checked 2^o(n)-time algorithm for 3-SAT, which would falsify the Exponential Time Hypothesis, alongside claimed Yang-Mills and Kakeya results, with academic shepherding as the verification mechanism and a prior 722-paper release that saw three withdrawals as the reason for skepticism.