On November 15, 2026, OpenAI launched a science blog covering findings, practical methods and researchers' firsthand experiences. Its opening guest post describes Nathaniel Craig's collaboration with GPT-6 Astra on a candidate lattice definition of the Standard Model.
A new account of an older paper
The underlying paper was submitted on October 6, 2026. The November event is a new publication and account of the research process. In his guest post, Craig says he supplied the refinement strategy and Astra helped develop the construction and proof. He also credits discussions with human experts.
The checks that remain open
Craig calls the construction a candidate requiring further expert scrutiny. Practical simulations remain unfinished, and reflection positivity at finite lattice spacing is still under examination. This digest has not independently checked the proof.
A proposed evidence ledger
Original teaching example, not an executed experiment: give an agent a synthetic mathematical claim and require a four-column ledger: claim, assumptions, supporting artifact and unresolved objection. When a proposed lemma changes, retain the rejected version beside the replacement. Have a separate checker inspect a small counterexample search, while a person decides whether that search addresses the actual assumptions.
The exercise should finish with an explicit list of what remains unproved. A persuasive explanation should not close an unresolved objection. Keeping the research artifact separate from the later narrative also makes it possible to track which claims changed and which checks still need an answer.
Sources accessed November 16, 2026. Views are my own and do not represent any employer, past or present.
Sources and evidence
The new event is a methods-focused publication launch. Keep the original paper date, human contributions and unresolved checks visible when discussing the candidate construction.