OpenAI’s October Science Release: Follow the Research Artifact Trail
A science announcement can point readers toward several different kinds of material: a research claim, a manuscript, a formalization and a later explanatory writeup. For OpenAI’s October 26, 2026 release, the practical starting point is to identify which material is public for each result.
Three release paths, not one package
OpenAI describes its model-produced manuscripts as raw research data and sets out different release arrangements. For the mathematics result on local smoothing, it releases model output for researchers to explain. For the computer-science result concerning the Exponential Time Hypothesis, it releases materials alongside an external collaboration. For the physics work, it defers release of materials to an academic group, which is initially preparing its own writeup.
OpenAI’s October 26 announcement: release arrangements and collaboration principles
The announcement says collaborators participate independently and may publish their conclusions. OpenAI also distinguishes recognition for explanatory work from credit for the original model-produced discovery. Those are the stated collaboration arrangements; this article does not establish scientific acceptance of the underlying results.
Start from the exact source link
Choose the result you want to follow and open its links from the announcement. Record the announcement date, artifact title, repository path and date you accessed it. Keep those details together so you do not accidentally attach commentary about one result to a different manuscript.
The linked computer-science manuscript lands at a GitHub PDF page whose folder names October 25. Its accompanying formalization link lands at lean/docs/378.md. These are different resources with different jobs. The PDF page exposed file metadata during this check, while the scope document’s text was readable.
Linked ETH manuscript: GitHub file page
Linked ETH formalization: scope document
Read the scope page before equating artifacts
The scope document separately describes what Lean proves and a sharper bound in the October 25 manuscript. It also links a comparator statement named SubexponentialSAT.lean. Readers should preserve those distinctions rather than treating every mathematical expression in the paper and formalization as an interchangeable statement.
For a reading notebook, create separate entries for the manuscript’s claim, the scope description, the formal statement and your own unanswered questions. A link between them is a route for investigation. It does not mean that opening one file has completed the work of understanding or checking the others.
Keep access and verification in separate columns
Mark whether you located a file, read its contents, followed its dependencies or reproduced a check. These are useful progress markers for a study group. A repository page that loads is evidence of access to that page; it is not a record that you ran a proof checker.
If a preview stays blank, record the limitation before summarizing what the file supposedly contains. If a later version appears, retain which version your notes describe. When sharing a reading list, distinguish source material from an explainer and from your own interpretation.
Follow the collaboration’s next output
For material that is being held for a group’s writeup, follow the stated release channel rather than assuming that the public links for another result supply the missing artifact. Date any reading guide so that a later release can be added without rewriting the history of what was available.
This is an artifact-navigation guide. The PDF proofs were not read in full, formalizations were not executed, and the mathematics, complexity-theory and physics claims were not independently validated here.
