Dominic Chiappe · People, capability & transformation

Thinking about how organisations perform in an AI-enabled world

AyEye Today ·

The proof trail is visible. The physics is not settled

Nathaniel Craig says GPT-6 Astra turned his refinement strategy into the detailed proof behind a candidate lattice definition of the Standard Model. The published artefacts make the work unusually inspectable, but they do not settle the result, its physical adequacy or the exact division of labour between researcher and model.

A dense ultramarine structure sweeps across broken lattice marks towards a small yellow flare on scraped warm-grey ground.
Original artwork: The Proof Between Scales. Chiappe × OpenAI.

On Sunday, OpenAI published an account by physicist Nathaniel Craig of work that tries to close a foundational gap in particle physics. Craig says he had a strategy for defining the Standard Model on a spacetime lattice but had not proved that it worked in every required case. GPT-6 Astra, he writes, suggested how to move force fields onto successively finer grids, showed that the comparisons converge and took the lead in developing the detailed proof.

The underlying preprint is not new. It appeared on arXiv on 6 October with Craig as its sole listed author. What changed on 15 November was the public account of the model's contribution, accompanied by an unusually extensive trail: a 120-page paper and supplement, a mathematical companion, a Lean formalisation, verification records, an analogous two-dimensional construction and a second route to the same measure.

That trail is the reason this clears the significance threshold. It gives outsiders more than a claim that an AI system helped. It exposes several places where the mathematics can be checked. It also shows why the headline cannot simply be “AI defines the Standard Model”. The paper calls the result a candidate. Craig says it is an answer in principle rather than in practice. Reflection positivity at finite lattice spacing is still under investigation, practical simulation remains future work and the interacting continuum theory remains open.

The right unit of judgement is therefore not one verdict. It is a chain of claims, each with its own evidence.

What problem is the construction trying to solve?

The Standard Model describes the known elementary particles and three fundamental forces with extraordinary experimental success. Yet many calculations begin by pretending interactions are switched off, solving that simpler problem, then adding the interactions back as progressively smaller corrections. That method, perturbation theory, works brilliantly when the series behaves. It is not a complete definition of an interacting theory in regimes where the interactions cannot be treated as small.

A lattice offers another route. Replace continuous spacetime with a grid, keep interactions present from the start and calculate on a finite set of points. Lattice methods have been powerful for the strong force. Extending them to the whole Standard Model is harder because the weak force treats left-handed and right-handed particles differently.

The difficulty is not a missing numerical constant. Straightforward lattice rules tend to produce extra partners, a problem known as doubling, while more sophisticated overlap-fermion methods move the problem into the phase of the integration measure. That phase has to vary smoothly, locally and consistently as the gauge field changes. Earlier work supplied important parts of the machinery, but the non-abelian case needed by the Standard Model remained open.

Craig's refinement strategy compares the allowed fermion states on a sequence of finer auxiliary grids. The differences between neighbouring scales are added to construct the phase rule on the original grid. The paper claims that the resulting current has the locality and global properties needed to reconstruct a smooth, gauge-invariant fermion measure for a class of chiral gauge theories that includes the Standard Model.

That is a precise mathematical target. It is also narrower than proving that the resulting lattice theory has every physical property one wants.

A worked audit: four claims inside one headline

Suppose a research director receives this briefing: “An AI system has proved a nonperturbative definition of the Standard Model.” Before acting on it, the director should split it into four claims.

First is the construction claim: the paper defines a gauge-invariant measure for overlap chiral fermions on the stated admissible gauge fields and satisfies the reconstruction criteria it invokes. The long paper, mathematical companion and formal code provide material for checking that claim.

Second is the correspondence claim: the Lean statements, definitions and assumptions faithfully represent the essential mathematics in the paper. The paper reports that all sixteen Lean modules compiled, kernel replay accepted the proofs and no unproved placeholders were used. Those checks establish the formal statements under their recorded assumptions. The paper itself warns that AI agents reviewed the correspondence between the formalisation and mathematical companion. A successful kernel replay cannot certify a mistaken transcription of the physical problem.

Third is the physics claim: the finite-spacing construction has the properties required to describe a unitary quantum theory and connects appropriately to continuum physics. This is where reflection positivity matters. Craig explicitly says that he is still investigating it. The paper also says the interacting continuum theory remains open.

Fourth is the contribution claim: Astra made the specific intellectual contribution described. The evidence here comes from Craig and the producer. The acknowledgements say the proof was Astra-led, that the manuscript contains extensive rewrites of Astra-composed proof text, and that Astra carried out tests proposed by Craig and colleagues. That is unusually candid provenance. It is not an independently reconstructed record of every prompt, suggestion, correction and rejected route.

The director can therefore make a bounded judgement: there is a serious, inspectable candidate result with a substantive model contribution reported by the researcher. Independent mathematical review, physical scrutiny and contribution analysis remain separate jobs.

Human review is evidence, not a transferable endorsement

Craig names three lattice field theorists who looked over the proof and asked probing questions. The preprint thanks Maarten Golterman and Yigal Shamir for sustained input and Andreas Kronfeld for expertise and comments. If Craig's account is accurate, it shows that the construction did not move directly from model output to publication without specialist challenge.

But the public account does not provide a signed statement from those experts that they endorse every theorem, the physical interpretation or the attribution of work. Their questions raised Craig's confidence. That is producer-reported scrutiny, not completed independent validation.

The distinction is not pedantry. Expertise can improve a proof while leaving its final status open. Reviewers can identify errors, ask for tests or help refine assumptions without becoming co-authors or guarantors. A responsible summary should preserve exactly what is evidenced: named experts engaged with the work; Craig reports that their questions were valuable; broader scientific dialogue is still required.

This also explains why the single-author byline should not be read as a complete map of intellectual labour. Authorship follows institutional and disciplinary rules. The acknowledgements provide a richer provenance statement than the byline alone. Even so, contribution reporting for model-assisted science needs a durable audit trail if readers are expected to distinguish brainstorming, proof search, formalisation, rewriting and final judgement.

What Lean checks, and what it deliberately leaves outside

Formal verification narrows one important source of uncertainty. A proof assistant can check whether a formal conclusion follows from the definitions, assumptions and prior results encoded in its environment. Replaying the result in Lean's kernel is stronger than saying the algebra looks plausible.

The paper does not treat that check as the final word. Its verification section says there is more that can go wrong in physics than a formally stated conclusion failing to follow from specified assumptions. It therefore adds a two-dimensional analogue with explicit calculations and a second construction that agrees up to the local counterterms the paper describes.

Those are useful controls because they fail differently. The formalisation tests deductive structure. The lower-dimensional model tests behaviour in a setting that can be examined more directly. The alternative construction asks whether a different route lands at an equivalent measure. None alone proves that the candidate has every physical property required at finite spacing or in the continuum.

This is the central practical lesson. Verification evidence should be attached to the layer it checks. “Lean passed” is evidence about the formal proof. It is not a universal seal on the physics, the exposition or the model's contribution.

Revisited: this is not the same proof story as 8 November

AyEye examined a different AI-generated proof on 8 November. That edition separated the formal statement, proof term, human explanation and downstream consequence. The same boundaries still matter here, but repeating them would miss the update.

This case adds two harder questions. The first is contribution. Craig describes a collaboration in which he supplied the strategy, Astra developed much of the proof, humans proposed tests and specialists challenged the construction. That is neither ordinary tool use nor a clean substitution of machine for researcher. It needs provenance that can survive beyond a launch essay.

The second is physical adequacy. In pure mathematics, statement matching and proof replay can bring a theorem close to settlement. Here a formally correct construction can still lack a property needed to represent the intended quantum theory. The verification ledger must therefore include unresolved physical criteria, not only proof status.

The advance is not that uncertainty has disappeared. It is that the artefacts make several uncertainties local and testable. An independent group can replay the formal code, inspect the assumptions, examine reflection positivity, compare the two constructions or attempt a practical implementation. Each successful check should raise confidence only where it reaches.

The counter-case: a polished audit trail can still be producer-shaped

OpenAI benefits when a visiting scientist attributes a major proof contribution to its newest model. The company hosts the explanatory post. The researcher is affiliated with OpenAI as well as UC Santa Barbara. The public evidence about the collaboration was selected and narrated by participants in the work.

That does not make the account false. It does mean the provenance claim has a different evidential status from the mathematical files. A sceptic could reasonably ask for timestamped development records, version history or a reproducible account of how the model changed the proof. Such records might reveal that Astra supplied the key construction. They might instead show that its largest contribution was elaboration, search or rewriting around a human insight.

The counter-case for the physics is equally concrete. A correct measure on the admitted field domain may still fail reflection positivity or prove too restrictive for useful simulations. If so, the work could remain mathematically interesting without delivering the claimed route to a satisfactory nonperturbative Standard Model.

The falsifier is not a vague demand that every physicist agree. It is specific. The strong reading weakens if independent reviewers find a fatal gap in the reconstruction, if the formal statements do not match the paper's essential claims, if reflection positivity cannot be established in the needed setting, or if contribution records do not support the reported Astra-led proof.

What would change confidence next

Four developments would materially strengthen the result:

  1. independent reproduction of the Lean build and a human comparison between the formal statements and the paper;
  2. specialist review of the locality, global integrability and anomaly arguments under the stated admissibility conditions;
  3. a resolution of reflection positivity, plus a clearer route from the construction to practical lattice simulations; and
  4. a versioned contribution record showing which proof steps originated with Craig, Astra and the human reviewers.

The paper has already done something valuable by publishing more of the trail than a conventional model-performance announcement. It makes the candidate proof open to targeted challenge. That is a better scientific posture than asking readers to trust a benchmark or a summary.

The appropriate conclusion is therefore strong but bounded. A physicist reports that GPT-6 Astra led the detailed proof of a substantial candidate construction, and the released materials give outsiders several ways to test it. The mathematical and physical result is not settled. The exact contribution split is not independently established. Those are not reasons to ignore the work. They are the next items on its verification ledger.

Sources and method

This edition treats Craig's OpenAI guest post and his arXiv preprint as producer and author evidence. The preprint is the primary technical source for the construction, formal checks, caveats and acknowledgements. The public account of Astra's contribution comes from Craig and OpenAI; no independent reconstruction of the collaboration was available in the reviewed window. A secondary report was used to check chronology and locate the disclosed review trail, not to validate the mathematics.

Review window: 14 November 2026 06:46 to 16 November 2026 06:46 Europe/London.

Chiappe × OpenAI. Primary-source review and research with OpenAI; editorial argument, evidence distinctions and final judgement by LecXie.