Axiom Math wants to keep AI slop out of mathematics

Axiom Math announced on Wednesday that its fleet of AI agents has formally verified the Stacks Project, an open-source mathematics reference started in 2005 by Columbia mathematician Johan de Jong and built up by contributors since. The company, which describes itself as a $1.6B startup, says the work translates roughly 7,654 pages of mathematics into about 16 million lines of computer-checked code. The Stacks Project is widely regarded as one of the largest bodies of mathematics assembled by humans; Columbia mathematician Peter Woit has called it comparable to the famous Bourbaki books, the 20th-century series that rebuilt mathematics from its foundations.
Formal verification is the technique at the center of the announcement. A proof is rewritten as code in a programming language called Lean, and a program known as a proof assistant checks that code line by line. If the proof passes the checker, its logical soundness is established at the level of the formal system, independent of whether a human reader finds the argument convincing. The same broad idea has been used in industry to confirm chip designs and, in France, to catch bugs in software that calculates family benefits. In practice, the hard part is not checking a single theorem but preserving the web of dependencies between results.
Theorems and definitions in Stacks build on one another, so a formalized version has to keep every connection intact across the full codebase. Axiom's CEO, Carina Hong, framed that tension directly: large and delicate things rarely go together, and a foundational collection of results needs its theorems to stay consistent and its definitions not to contradict each other. The company said it was initially unsure the project would succeed for exactly this reason.
To manage the scale, Axiom gave its agents a loose hierarchy rather than dictating every step. Ashvin Swaminathan, a member of the technical staff, said the agents found tasks for themselves and identified work they could do in parallel. The company also trained its own models for mathematical work, an approach Hong said most other verification startups have not taken. At first an in-house Lean expert taught the agents a human sense of taste, and over time, according to Hong, the agents picked up that judgment on their own.
The result addresses a notable gap. Swaminathan described formalized mathematics as having a large black hole covering algebraic geometry, meaning researchers who wanted to check new work in that area with Lean had little to build on. With algebraic geometry now represented, that foundation exists. Axiom argues each verified result becomes a reusable building block, so later proofs can lean on checked mathematics instead of starting over. The UCLA mathematician and Fields Medal winner Terence Tao has made the general point that the more formalized proofs exist, the more formalized proofs become possible.
Axiom was founded in March 2025 by Hong, who left a Stanford mathematics Ph.D. program and is 25. In March the company raised a $200 million Series A led by Menlo Ventures. While mathematicians are the immediate audience, the company's larger bet is on software. Developers have long struggled to guarantee code is free of bugs and security holes, and that problem grows as AI agents write more code. Axiom's argument is that verified mathematics is the place to start building such foundations and that the skill transfers: agents can read the Lean code, understand it and build new infrastructure on top.
This suggests the strategic claim is really about reuse. A formal library is only valuable if downstream tools and agents can depend on it without re-checking everything, which makes coverage of a field like algebraic geometry more consequential than the page count. The trade-off is that formal verification demands significant upfront effort and specialized expertise in Lean, and a verified library reflects the definitions someone chose to encode. The announcement establishes that the Stacks material was formalized; it does not establish that every future proof built on it will be correct, since errors can still enter at the modeling stage before the checker runs.
Hong's broader thesis is that agents will need a well-built, reusable body of knowledge to avoid producing low-quality output. She compared bad engineering choices at the start to a "slop drop" that makes it hard for agents to extend the boundary later, and she openly challenged coding companies to attempt formalization at this scale.
Why it matters: Researchers working in algebraic geometry can now build Lean proofs on a large verified base instead of starting from scratch, which lowers the cost of checking new results. For developers and product teams, the more relevant signal is that Axiom is pitching verified foundations as infrastructure for AI-written code, an inference about where the company intends to go rather than a shipped developer product. Teams relying on agents to generate code should expect growing interest in proof-checked foundations, while recognizing that formalization guarantees only what has been encoded, not the correctness of the model or specification behind it.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
As AI floods math with unchecked proofs, Axiom Math is betting on verification. The startup’s AI agents just formally verified the Stacks Project, one of the field’s biggest libraries.