Physicist Nathaniel Craig says GPT-6 Astra took charge of the detailed proof behind a candidate way of defining the Standard Model on a grid, and he describes the result as unfinished. Craig heads the particle theory group at the University of California, Santa Barbara. He wrote about the work on 15 November in the first guest post on OpenAI’s new science blog. He is also a visiting scientist at OpenAI, which matters when weighing who is vouching for the model.

The Standard Model is physics’ working theory of the known elementary particles and of electromagnetism, the weak force and the strong force. Experiments agree with it over eighteen orders of magnitude. Yet it still lacks a definition that avoids perturbation theory, the method of switching interactions off, solving the simplified problem, and adding the interactions back as small corrections. That method pins down how magnetic an electron is, with an error of a few parts per trillion. It fails where interactions dominate: most of a proton’s mass comes from quarks and gluons moving and interacting, and with the strong force turned off the proton disintegrates.

The standard workaround is lattice field theory, which swaps smooth spacetime for a grid. A finite grid gives a rigorous definition and something a computer can calculate, and it has already explained the proton’s mass from first principles. The Standard Model resists it because the weak force treats left-handed and right-handed particles differently. Put such a theory on a grid naively and each left-handed particle arrives with an unwanted right-handed twin, a problem physicists call doubling.

Craig’s analogy is a mirror. An ordinary mirror turns light spinning one way into light spinning the other way, while a chiral mirror, like the spiral films in some scarab beetle shells, returns it spinning the original way. Decades of work by Paul Ginsparg, Ken Wilson, Rajamani Narayanan, Herbert Neuberger, Martin Luscher and Yoshio Kikukawa supplied most of the machinery. One ingredient, a correction to the quantum phase, stayed unsolved for the full Standard Model for over 25 years.

His proposal re-expresses the force fields across a sequence of progressively finer grids. Comparing the allowed particle states from one grid to the next yields a slice of the needed phase fix, and summing the slices produces the rule for the original grid. According to Craig, the proof shows the correction is right around every allowed loop, leans mostly on neighboring fields, and agrees across different ways of describing the same fields. It holds only if field strengths at the grid scale stay capped.

Craig says he had the strategy but could not show it worked in every case. He calls GPT-6 Astra essential to turning the refinement idea into a working construction: the model proposed how to transfer force fields to finer grids, then showed that the comparisons converge, and led the detailed proof. The work took several rounds with the model plus conversations with external specialists. Lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir read the proof and pressed him with hard questions.

The caveats come from Craig himself, and they are heavy. The result is a candidate definition that might settle a puzzle unresolved for half a century, yet he calls it “much more a beginning than an end.” Experts still have to decide whether the prescription has every property it needs. He also calls it “far more an answer in principle than an answer in practice,” since realistic lattice simulations will take much more work. He is now testing a property called reflection positivity at finite grid spacing, and the construction leaves the continuous-spacetime theory undefined. The post offers no independent verification, and the account of the model’s contribution is Craig’s own.

For the industry, the claim reaches past a benchmark score. A lab that can point to a model leading a proof in a field with a 50-year gap is selling research capacity, and the buyers are universities and national labs rather than chatbot users. Whether that pitch holds depends on what Golterman, Kronfeld, Shamir and their peers conclude, and physicists will wait for that verdict before treating the proof as settled.

Based on “A step toward defining the Standard Model on a lattice,” a guest post by Nathaniel Craig on the OpenAI science blog, published 15 November 2026.