What happened
The AxiomMath/Mordell repository describes a Lean 4 formalization of Faltings’ theorem, following Enrico Bombieri’s 1990 proof, and a theorem of Paul Vojta used in that argument. The theorem states that for a smooth, projective, geometrically integral curve over a number field, when the genus is at least two, the set of rational points is finite. The repository also gives an equivalent formulation for function fields of one variable: there are finitely many degree-one places.
The formalization is built on Mathlib and the Stacks Project in Lean. The repository says the library imports about 430 modules in Mordell, which were written by AI agents. It also says the formalization proves the LeanEval problem’s statement verbatim, with its definitions, and includes a second statement about the finiteness of the poles of a function-field element. The repository identifies specific formalized parts of Bombieri’s paper with Lean attributes carrying labels for the corresponding items.
The README sketches how the mathematical argument is represented. The Mordell–Weil theorem gives finite generation for the Jacobian’s rational points; the Néron–Tate height equips the associated real vector space with a Euclidean structure; and Bombieri’s form of Vojta’s inequality constrains the geometry of points of large norm. Northcott’s theorem bounds points of small norm. The code’s theorem declarations are checked by Lean’s kernel, and the README says there is no sorry, native_decide, or added axiom. It names the standard axioms on which proofs depend: propext, Classical.choice, and Quot.sound.
The project is published under the Apache 2.0 license. Its README provides build commands (lake exe cache get and lake build) and warns that the first build takes a long time and a lot of memory because Stacks Project modules are compiled locally. The GitHub page showed six stars and two forks when captured; those counts are snapshots, not measures of correctness or mathematical adoption.
What the archive adds
New Runtime’s earlier coverage on November 5, 2026 described Axiom’s Lean formalization of the Stacks Project, around 16 million lines in a reusable library. This is a new result built on that foundation, not an independent review of the earlier library. An earlier New Runtime note on November 16, 2026 also raised the distinction between a claim about formalized mathematics and public code that lets others check it; this repository makes the code available while still disclosing that human review has not happened.
Trend signal
This supports the scientific-agents-need-artifact-evidence pattern: the result is delivered as a concrete artifact whose declarations and axiomatic dependencies can be inspected and checked, rather than only as a claim of success. The caveat matters just as much as the formal verification: kernel checking does not establish that humans have reviewed the library’s design, coverage, or correspondence to the intended mathematics.
Why now
The repository’s stated dependence on the Stacks Project in Lean gives this result a specific prerequisite: substantial formalized foundations for algebraic geometry. The announcement shows a further theorem being built on that base, extending the body of mathematics represented in Lean rather than presenting a standalone proof detached from reusable infrastructure.
Why it matters
For teams building AI systems that produce proofs or other high-stakes technical artifacts, the repository illustrates a useful separation of responsibilities. AI agents can generate a large body of formal code, while a proof assistant’s kernel checks whether each accepted proof follows from the formal definitions and axioms. That is stronger than trusting a model’s natural-language explanation, and the explicit axiom list makes part of the trust boundary inspectable.
But machine checking is not the same as independent mathematical review. A checked artifact can still require scrutiny of its formalization choices, dependencies, and whether the encoded result matches the intended theorem. The project’s own disclosure—that its AI-written code has not been reviewed by humans—should travel with any account of the result. Builders should treat proof assistants as verification infrastructure, not as a reason to skip review of what has been formalized.
Limits and uncertainty
The repository says every proof is kernel-checked and names the standard axioms used, but the code has not yet been reviewed by humans. That leaves open how independent experts will assess the formalization’s structure and correspondence to the intended argument. The README describes the result and its mathematical dependencies; it does not report a human audit, an independent reproduction, or an evaluation comparing the AI agents’ work with a human-authored formalization. The GitHub star and fork counts do not measure any of those things.
Skill to build now
When adopting a machine-checked result, inspect the theorem declaration, its imported library, and the proof assistant’s reported axioms before relying on a headline claim. In Lean, the repository demonstrates #print axioms as a direct way to expose a declaration’s dependencies. For a large AI-written library, pair that check with a human review plan: identify the definitions that encode the mathematical claim and the dependencies most consequential to it.
Reader takeaway
Axiom has published a Lean-checked formalization of Faltings’ theorem on top of its Stacks Project formalization. The kernel-checked proof and disclosed axioms make the result inspectable; the authors’ statement that the AI-written code has not had human review is an equally important part of the story.