The Riemann Proof Nobody Can Read Just Passed Every Checker
Twenty-four hours ago we called Levent Alpöge's Lean proof of the Riemann Hypothesis "a claim, not a verified result." That line is now out of date. On Saturday at 13:55 UTC the Lean and Mathlib communities, with the Lean FRO and the Mathlib Initiative endorsing, published a joint statement: "We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct," and "we hereby certify that both the statement and the proof of the Riemann Hypothesis are correct." It went straight to the top of Hacker News with 679 points. Statement: https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
How do you certify 17 million lines nobody will ever read? You stop trusting any single thing. The statement being proved is Mathlib's own RiemannHypothesis definition, vetted by Lean and number theory people, so nobody snuck in an easier theorem. Dozens of people compiled it and got the same raw export file. A separate challenge file that imports only Mathlib was compared against Alpöge's theorem with two independent tools. Then the export went through Lean's official kernel plus con-leche, con-ron and third-party checkers written in other languages. Two of those checkers come with a formal proof, relative to set theory, that they cannot accept a false proof. And every check ran on an inert export, out of reach of any malicious code hiding in the proof sources. Short version: nobody reads the proof, so they check the checkers.
The best detail is on Andrew Sutherland's page at MIT: "checks run with Claude Code." So a proof made with Claude was audited by a mathematician driving Claude. Every one of 175,187 modules compiles with zero errors. The only axioms are propext, Classical.choice and Quot.sound. Mathlib's .olean files are byte-identical to the official cache. nanoda, an independent kernel written in Rust, accepts the full export in 27 minutes 42 seconds on 80 threads. con-leche accepts 1,468,418 declarations in 3 minutes 15 seconds. As a bonus, the team plugged it into a separate Robin formalization and got Robin's inequality for n > 5040 for free. Logs: https://math.mit.edu/~drew/grh/
Now the fine print, because it matters. The "generalized" part checked here covers Dirichlet L-functions only. Dedekind, Hecke and general automorphic L-functions "are not among the statements that were checked." Lean's default and paranoid kernel runs plus the end-to-end comparator were still listed as running when we last read the page. Someone in Terence Tao's comments asked whether Lean's type theory, which is stronger than ZFC, changes what has been proved, and no primary source has answered yet. Nobody has certified the Navier-Stokes package, which still has a missing certificate and an unrebuilt patch. And Alpöge works at Anthropic, which is the company whose model helped and whose IPO just closed.
What the statement leaves open is the interesting part. The community closes by looking forward to work "that will help explain the new ideas in this proof." So it's true, and nobody understands it yet. For a century, being correct and being understood arrived together, because a human proof was the explanation. Now they've split, and checking may turn out to be the cheap half. If that holds, the scarce job in math stops being finding proofs and becomes explaining them.
This article was written with Claude, an Anthropic model.
← Back to all articles
How do you certify 17 million lines nobody will ever read? You stop trusting any single thing. The statement being proved is Mathlib's own RiemannHypothesis definition, vetted by Lean and number theory people, so nobody snuck in an easier theorem. Dozens of people compiled it and got the same raw export file. A separate challenge file that imports only Mathlib was compared against Alpöge's theorem with two independent tools. Then the export went through Lean's official kernel plus con-leche, con-ron and third-party checkers written in other languages. Two of those checkers come with a formal proof, relative to set theory, that they cannot accept a false proof. And every check ran on an inert export, out of reach of any malicious code hiding in the proof sources. Short version: nobody reads the proof, so they check the checkers.
The best detail is on Andrew Sutherland's page at MIT: "checks run with Claude Code." So a proof made with Claude was audited by a mathematician driving Claude. Every one of 175,187 modules compiles with zero errors. The only axioms are propext, Classical.choice and Quot.sound. Mathlib's .olean files are byte-identical to the official cache. nanoda, an independent kernel written in Rust, accepts the full export in 27 minutes 42 seconds on 80 threads. con-leche accepts 1,468,418 declarations in 3 minutes 15 seconds. As a bonus, the team plugged it into a separate Robin formalization and got Robin's inequality for n > 5040 for free. Logs: https://math.mit.edu/~drew/grh/
Now the fine print, because it matters. The "generalized" part checked here covers Dirichlet L-functions only. Dedekind, Hecke and general automorphic L-functions "are not among the statements that were checked." Lean's default and paranoid kernel runs plus the end-to-end comparator were still listed as running when we last read the page. Someone in Terence Tao's comments asked whether Lean's type theory, which is stronger than ZFC, changes what has been proved, and no primary source has answered yet. Nobody has certified the Navier-Stokes package, which still has a missing certificate and an unrebuilt patch. And Alpöge works at Anthropic, which is the company whose model helped and whose IPO just closed.
What the statement leaves open is the interesting part. The community closes by looking forward to work "that will help explain the new ideas in this proof." So it's true, and nobody understands it yet. For a century, being correct and being understood arrived together, because a human proof was the explanation. Now they've split, and checking may turn out to be the cheap half. If that holds, the scarce job in math stops being finding proofs and becomes explaining them.
This article was written with Claude, an Anthropic model.
Comments