RESEARCH · FOCUS

AxiomMath announces the formalization of the Stacks Project in Lean

AxiomMath says it has formally verified the Stacks Project using its AI agents. The announcement concerns a vast mathematical library translated into Lean, with the ambition of providing reusable foundations for other proofs and, eventually, for software.

AxiomMath announces the formalization of the Stacks Project in Lean — An original illustration of a mathematical manuscript and its reasoning.
An original illustration of a mathematical manuscript and its reasoning.DailySynth illustration

AxiomMath says it has formally verified the Stacks Project, a reference library in mathematics. In an article published on November 5, TechCrunch reports that the company’s AI agents translated its roughly 7,654 pages into around 16 million lines of computer-verified code. The announcement was made on Wednesday, according to the outlet.

The work relies on Lean, a language in which proofs are reformulated as code and then checked by a proof assistant. For a library of this scale, the challenge extends beyond verifying each theorem individually: definitions and results depend on one another. AxiomMath explains that it had to preserve these relationships throughout the translation.

AxiomMath announces the formalization of the Stacks Project in Lean — A conceptual diagram: a proposition, its proof and its verification.
A conceptual diagram: a proposition, its proof and its verification.DailySynth illustration

The company presents this formalization as a reusable foundation, particularly in algebraic geometry. According to its team, researchers seeking to verify new work in this field previously had few foundations available in Lean. Its agents worked in parallel within a hierarchical structure that gave them latitude to identify their tasks.

The potential value lies in accumulating checked results on which other proofs could build. AxiomMath also wants to transfer this approach to software. That prospect remains an ambition: the reported announcement concerns mathematical formalization and does not, on its own, establish that agents can produce software free of defects.

This article is part of the edition of November 5, 2026.Read the full edition ↗

Also read.

FlashProve: Anthropic explores proof of AI model identity

Anthropic is working on FlashProve, a system intended to cryptographically verify which model produced a response. According to TechCrunch, specifications and code are available, but the research remains at an intermediate stage and no deployment is planned.

Nov 4 · 2 min↗

OpenAI's scientific results: access to evidence sparks debate

According to The Verge, researchers challenge the arrangements for releasing results announced by OpenAI in mathematics, computer science and physics. Their potential scientific significance remains distinct from their validation, which is complicated by differing levels of access to the work.

Oct 28 · 2 min↗

OpenAI submits its scientific announcements for researchers’ scrutiny

OpenAI announces results in mathematics, theoretical computer science and physics. The process described by The Verge combines publication of work with expert groups, while advisers challenge some early-access practices. These announcements remain proposals to be examined.

Oct 27 · 1 min↗