What was actually done

The classification is a statement about the building blocks of finite groups: every finite simple group is cyclic of prime order, alternating, one of sixteen families of Lie type, or one of 26 sporadic groups. Humans proved it across decades, finishing the first-generation proof in 2004 when the quasithin case was finally published. The first-generation classification spans roughly 15,000 pages across hundreds of publications. Independently auditing every inference across that literature is an enormous undertaking, especially when the arguments were not originally written for machine verification. That is the job Axiom gave its agents.

Formalizing is translation with a strict editor. Each published argument had to be restated in Lean precisely enough for the kernel, a small trusted checker, to accept every inference. An agent can propose an incorrect argument, but Lean's kernel will reject a proof term that does not establish its stated formal goal under the accepted axioms. That check is powerful, although it does not by itself guarantee that every formal statement perfectly matches the intended mathematics. Lean's kernel checks the logical inferences in the submitted proof terms, reducing the need for humans to verify each line manually. That mechanical check was essential to coordinating work at this scale, but engineers and mathematicians still directed the project, managed dependencies and intervened when parts of the formalization had to be rebuilt.

The gap in the 1986 proof

The strongest result in Axiom's account is not the line count. While formalizing John Walter's 1986 treatment of the Unbalanced Group Theorem, a load-bearing piece of the classification, the agents found explicit counterexamples to an assertion the argument relied on: an operation Walter assumed kept certain groups inside a specified family can produce a group the family excludes. Axiom's worked example starts from PSL2(49) and lands on PSL2(7), excluded because 7 is a Mersenne prime.

Two qualifications matter, and Axiom states both. The counterexamples did not refute the theorem, or even the conclusion of the proposition in question. They left one inference unjustified, which for a machine-checked proof is enough to stop the line. Closing it took a different argument, built on Ronald Solomon's 1976 theory of maximal 2-components, whose key step Axiom describes as a new result proved by its AxiomProver system.

Then comes the detail that should keep the story honest. After the formalization, Axiom's own mathematicians found a shorter deduction, inside the original first-generation methods, that reaches the same conclusion without the new theorem. The machine's contribution was real: it found the hole and produced a working repair. It was not the only repair available once humans looked at the hole it had found.

Reading the 43.1 million

The 43.1-million-line figure needs two qualifications. First, about 1.2 million of those lines came from Qiuzhen-CFSG, an earlier formalization project, and the whole development stands on Mathlib, a library built by hundreds of contributors. The 24 days produced the combined, complete codebase, not 43.1 million fresh lines. Second, "complete" is currently Axiom's claim. Axiom says the public repository is sorry-free and that a complete verification build takes less than 24 hours. No independent reproduction of that full build was identified in the sources reviewed for this article. The result is publicly available for checking, but Axiom's end-to-end verification claim has not yet been independently reproduced in this reporting.

Axiom, to its credit, frames the result the same way. Its account closes by assigning the classification to the hundred-plus mathematicians who proved it, and claims for itself the narrower work: making the proof mechanically checkable, documenting the obstructions, and formalizing arguments that resolve them. The errata list, 34 noteworthy items, is being compiled for publication with the classification's own surviving architects advising.

The practical read

The project illustrates an important advantage of agent workflows in formally checkable domains: proposed outputs can be tested against explicit constraints rather than accepted on the agent's authority. In Lean, a proof term either satisfies the stated formal goal under its accepted axioms or it does not. The CFSG project also shows the boundary. Axiom reports that its agents formalized the large proof corpus and surfaced hidden assumptions, including the gap in Walter's argument. Afterward, the company's mathematicians identified a shorter deduction using existing first-generation methods. The two routes show different contributions to the same result without establishing that one approach could not have found the other's argument.

Sources

  1. [1] Axiom Math, "Formalizing the Enormous Theorem," October 2026Read source
  2. [2] Axiom Math, project page "CFSG in Lean"Read source
  3. [3] Unite.AI, "Axiom Math Releases 43.1M-Line Lean Proof of Group Classification," October 27, 2026Read source
  4. [4] Defant and Ono, "The Unbalanced Group Theorem," Hexagon preprint server, version 1, October 27, 2026Read source
  5. [5] AxiomMath/CFSG repository, READMERead source
  6. [6] Wikipedia, "Classification of finite simple groups" (background on the first- and second-generation proofs)Read source