The Stacks Project formalized in Lean by directed AI agents
Axiom Math released a Lean formalization of the whole Stacks Project, the roughly 7,500-page reference work on algebraic geometry, at about 16 million lines with no sorry, and says its agents found dozens of statements in the original text needing repair, including a lemma on algebraization of coherent formal modules that the library marks as false as published.
- Lab
- Axiom Math
- Model
- AxiomProver agent swarms using an unnamed mix of models
- Field
- Mathematics
- Date
- 2026-11-04
What was found
The Stacks Project, started by Johan de Jong in 2005, is an open-source reference whose every item carries a permanent four-character tag. The release formalizes every tag with formalizable content on top of Mathlib, preserving the tag structure so each Lean declaration links back to the canonical text, and tagging declarations with the EGA and SGA items they prove. The library is split into two tiers: StacksCore, which a person has checked against the source and brought to Mathlib standards, and StacksFrontier, written by AI agents and not yet reviewed, with Lean's kernel checking every proof in both and only the three standard axioms permitted. Axiom says agents diagnosed the obstructions behind failed proof attempts and built explicit counterexamples, and the repository carries them as named files: StacksFrontier records that tag 0EJA is false as published because its proof identifies a dimension function with the dimension of a closure using a result that needs catenary, and gives a pinched Noetherian local domain as a counterexample, alongside principal-ideal and two-generator counterexamples to a canonical map on sections away from a closed subset.
Novelty check
This formalizes an existing reference text rather than proving a new theorem, and is graded as such. arXiv, 2026-11-09: abs:"Stacks Project" returned 11 records, none of them a formalization of it. Within this registry Axiom Math already holds 2026-10-26-cfsg-lean-axiom, released eight days earlier, which this entry parallels: that one was entered because the agents found a concrete counterexample to a step in one published route and a companion paper resolved it. The counterexample claim here was checked against the source: upstream tag 0EJA on stacks.math.columbia.edu is unchanged and carries no comments as of 2026-11-09, so the correction has not been taken up by the Stacks Project itself.
Caveats and known objections
Graded claimed, following this registry's practice for large formalizations nobody outside the authors has rebuilt: every check recorded is Axiom's own. The counterexamples that make this more than a translation exercise all sit in StacksFrontier, the tier Axiom itself describes as written by AI agents and not yet reviewed by a person, and Axiom has not said what fraction of the library reached StacksCore. Lean's kernel checking a counterexample establishes that the formal statement holds, not that the formal statement faithfully renders the Stacks Project item it is attached to, and faithfulness is enforced by Axiom's own multi-agent auditing. The library has about 63,000 modules and no compiled cache service yet, so a full rebuild takes a long time and a lot of memory and no independent replay was found. Johan de Jong asked on 2026-10-21 that AI-generated comments and pull requests not be sent to the upstream repository for now, and Axiom keeps its corrections in its own repository, so outside confirmation should not be expected soon. The models are not named. Autonomy is collaborative: agents wrote the Lean and constructed the counterexamples, while Axiom's mathematicians and engineers designed the agent hierarchy, the coordination protocols and the audits.
No independent check is recorded on this entry yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Claimed and Collaborative.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Community discussion
Graded claimed for verification and collaborative for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). The Stacks Project formalized in Lean by directed AI agents. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-11-04-stacks-project-lean-axiom
BibTeX
@misc{whataifound-axiommath-2026-axiom,
title = {The Stacks Project formalized in Lean by directed AI agents},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Axiom Math. Verification: Claimed. Autonomy: Collaborative.},
url = {https://whataifound.org/finding/2026-11-04-stacks-project-lean-axiom}
}