Axiom Math says its AxiomProver multi-agent system has produced a complete, computer-checkable Lean formalisation of the Classification of Finite Simple Groups. The project took 24 days and produced about 43.1 million lines of Lean and 2.05 million declarations. Along the way it flagged 34 notable deviations from the published proofs. The headline result is real and public, but the review is self-assessed so far. For anyone building agents, the more useful lesson is how the work was coordinated, not the line count.
What Axiom announced
According to TokenPost’s report on the AxiomProver formalisation, Axiom announced on 26 October that it had completed what it describes as the first complete formalisation of the result. The original proof spans about 15,000 pages and draws on more than 500 papers. More than 100 mathematicians contributed to it over several decades. The theorem says every finite simple group is a cyclic group of prime order, an alternating group, a member of one of 16 families of Lie type, or one of 26 sporadic groups.
AxiomProver splits the work across agents for decomposition, search, proof-writing and review. Kingy AI’s detailed breakdown stresses that this was human-guided. Agents wrote and revised formal proofs while the team handled mathematical obstacles and engineering infrastructure. The project also builds on earlier formal libraries, including Mathlib. Kingy notes that the 24-day clock does not include building those libraries or the decades of human mathematics behind them. Axiom has also not disclosed compute cost, hardware, its full model inventory or person-hours.
What is and isn’t verified yet
A Lean kernel check proves that the encoded deductions follow from the encoded assumptions. It does not prove that the formal statement captures the theorem mathematicians actually mean. TokenPost makes this point, and Kingy expands on it: a flawlessly checked proof of a weakened statement would settle nothing. Kingy also points out that Axiom’s own project metadata describes the review as self-assessed. Kingy has not reproduced the build itself.
Axiom’s CEO, Carina Hong, was asked by RuntimeWire about independent reproduction of the proof whether anyone outside the company had built and checked it. She answered: “Not yet. We are publishing the build artifacts to help and encourage external parties to independently reproduce it.” Axiom says a full build can take under 24 hours with its optimisations. That figure is also unverified.
The 34 deviations and the repair
The most interesting output may be what the agents found in the human literature. TokenPost reports that the team found a significant gap in one proof of the Unbalanced Group Theorem and wrote a new segment of proof to close it. It also classified 34 departures from the published record as notable deviations.
Axiom’s accompanying paper, dated 27 October, gives a counterexample to a specific closure assertion using an automorphism of PSL₂(49), then offers an alternative argument. Kingy is careful about scope here. A defective step in one route does not refute the Unbalanced Group Theorem or the classification itself. The paper’s host also does not conduct peer review.
Our read is that this is still a strong signal. Accepted, heavily cited work contained steps that did not survive mechanical checking at full scale. Human review bandwidth had not caught them in decades.
What it means for builders
The part worth copying is the architecture. Kingy describes the hard problems as coordination problems:
- definitions have to agree across files;
- one agent’s required assumptions must match what another agent has already proved;
- a change deep in the dependency tree ripples outward;
- discarded conditional proofs and repeated computation have to be managed.
That is the same shape as migrating a large legacy codebase, auditing a regulatory corpus or reconciling years of contracts. The difference is that Lean gives exact, cheap, binary feedback on every step. Most business domains have nothing like that.
So the honest takeaway is narrower than “agents can now do anything long-horizon.” Agents can sustain work across a huge dependency graph when there is a strict checker in the loop and humans handle the genuinely hard calls. Remove the checker and you are back to agents grading their own homework. That is exactly the failure mode the self-assessed label warns about.
What to do now
Look for the “CFSG problem” in your own operation. It is a body of work too large for people to review end to end, but with parts that can be checked mechanically. Typical candidates are reconciliations, schema migrations, policy-to-control mappings and test suites against specifications.
Build the verifier before you scale the agents. Then split roles the way Axiom did, with separate agents for decomposition, search, generation and review. Our comparison of LangGraph, CrewAI and custom orchestration covers the trade-offs in wiring that up.
For regulated outputs, the same pattern of machine checks plus a named human reviewer underpins sensible compliance reporting automation.
Finally, treat vendor claims the way Kingy treated this one. Ask what was independently reproduced, what the checker actually checks, and what the elapsed-time figure leaves out. For contrast with chat-assisted maths, see our note on Meta’s Muse Spark and five open maths problems.
Frequently Asked Questions
Has anyone independently verified Axiom's Lean proof?
Not according to Axiom's CEO. Asked by RuntimeWire, Carina Hong said "Not yet" and that the build artifacts are being published to encourage external reproduction. Kingy AI notes the project metadata marks the review as self-assessed.
Does the flaw Axiom found mean the classification theorem is wrong?
No. The counterexample targets a specific closure assertion in one published route to the Unbalanced Group Theorem, and Axiom offered a replacement argument. As Kingy AI points out, a defective step in one route does not refute the theorem or the classification.
Was this done by AI alone?
No. Axiom describes a human-guided effort: agents wrote and revised proofs, while mathematicians and engineers handled obstacles and infrastructure, and expert mathematicians oversaw review. It also built on existing libraries such as Mathlib.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Set up a free 30-minute call and we will map your first win together.
Get in touch
