What happened
- Axiom Math, a startup valued at US$ 1.6 billion, announced on Wednesday 4 November 2026 that it had formally verified the Stacks Project. TechCrunch reported the case on 5 November.
- The Stacks Project is an open reference started in 2005 by Columbia mathematician Johan de Jong. It gathers results in algebra, geometry and number theory, from college level up to the field’s advanced topics.
- Axiom’s fleet of AI agents translated its roughly 7,654 pages into about 16 million lines of code in Lean, a language in which a program called a proof assistant checks each proof line by line.
- The agents ran in parallel under a loose structure. According to Axiom’s Ashvin Swaminathan, there was a hierarchy, but the agents found on their own the tasks they could tackle at the same time.
Why it matters
- If a proof passes the proof assistant, it can be trusted to be logically sound. It is another way of checking mathematics, different from peer reading.
- The result targets a known gap. There is “this huge black hole” in what has been formalized, and it covers algebraic geometry, Swaminathan told TechCrunch.
- The company sees a use beyond mathematics: a verified base on which agents can write software with fewer bugs and security holes.
The number
16 million lines. That is the approximate size of the Lean code. The difficulty lay not only in each theorem, but in keeping intact the connections between definitions and results that rest on one another.
Context
Axiom was founded in March 2025 by Carina Hong, 25, who left a mathematics PhD at Stanford. In March this year it raised a US$ 200 million Series A led by Menlo Ventures. Hong told TechCrunch the team was unsure whether coding with agents would preserve the structure of Stacks, and that definitions must not contradict one another.
The company also trained its own models for mathematics. At first, an in-house Lean expert taught them judgment; over time, according to Hong, they picked it up on their own. That is a company claim that could not be checked against an outside review in the TechCrunch piece.
At Mamífero we have covered verification and work with agents. We reported that Claude tested the security of 123 open libraries with formal verification and that ThinkingBox rewards the agent for the database it leaves behind. Also that Recursive released Flint, a coding model for laptops and that GitHub Copilot added projects with several agents.
What’s next
- Researchers in algebraic geometry will have material for checking their new work in Lean, something that barely existed before, according to Axiom.
- Each verified result serves as a piece for future proofs, so that no one has to start from scratch.
- Outside mathematicians have yet to judge how useful this translation is for their daily work.
Bottom line
A library written over two decades by people now also exists as code a machine can check. Its value will be measured by how many new proofs rest on it.
Sources
- TechCrunch, “Axiom Math wants to keep AI slop out of mathematics”, 5 November 2026, link.
- Cover image: Tungsten, Public domain, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
