Riemann Hypothesis Proof Verified by Two Groups: What Changed
Key Takeaways
- Two organisations, the Lean Focused Research Organization and the Mathlib Initiative, announced on Saturday 21 November that they had independently verified Levent Alpöge's claimed proof of the Riemann hypothesis, saying they have "full confidence" it is correct.
- The verification is of the formal proof: around 17.4 million lines of Lean code that compile without errors and were rechecked by the Lean kernel using only the three standard axioms.
- Independent checkers written separately from Lean — nanoda, con-leche and lean4lean — all accepted the exported proof, and con-leche's compare mode matched all five headline theorems against the standard statement.
- The most delicate check, the comparator run, confirms the code proves the Riemann hypothesis as mathematicians actually state it, rather than a redefined or easier version — although reports differed on whether the full end-to-end run had finished.
- What has not changed: nobody has yet read and understood the whole mathematical argument, there is no human-readable proof, and the Clay Mathematics Institute still lists the problem as unsolved.
- A rival human-written claim to the generalised version, by Stanford researcher Alessandro Masullo, is undergoing major revision after mathematicians reported problems with it.
What happened on Saturday
On Friday 20 November, Levent Alpöge, a mathematician at Anthropic, posted a four-page note on his personal website saying: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He released the argument not as a conventional paper but as computer code written in Lean, a programming language in which a mathematical statement and the reasoning behind it can both be expressed precisely enough for a machine to check every step. The archive ran to 14 parts and roughly 17.4 million lines of code. Alpöge said the work with Claude had taken about six weeks, and that the same method also settles the Generalised Riemann Hypothesis for Dirichlet L-functions, a broader family of related functions.
Our earlier coverage of the claim, published yesterday, explained what was at stake and noted the one check that still remained: the "comparator run", the test that confirms the code does not redefine symbols or quietly prove something easier than the real Riemann hypothesis. That check, and the wider independent verification, is what changed on Saturday. As The Guardian reported, two groups verified the mathematician's claim that Claude helped prove the mathematics problem, and the announcement landed on Saturday afternoon, barely a day after the claim itself. If you are catching up, our first report covers the claim in full: Anthropic Riemann Hypothesis AI Proof.
Who the two groups are, and what they actually certified
The two groups named in the announcement are the Lean Focused Research Organization, which supports the development of the Lean programming language, and the Mathlib Initiative, which supports Mathlib, Lean's large shared library of formalised mathematics. Their joint statement, carried by Scientific American, was unambiguous: "We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement." The statement went further still: "We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct."
That wording deserves careful reading, because the word "certify" can easily be misunderstood. The Lean and Mathlib communities are the groups behind the software that checked the proof. Their certification covers the formal statement and the computer-checked proof. It is not a journal's verdict, it is not peer review in the traditional sense, and it is not a ruling from the Clay Mathematics Institute, which continues to list the Riemann hypothesis as unsolved and which administers the one-million-dollar Millennium Prize for the problem. USA Today made the same distinction, noting that the announcement "does not mean mathematicians have finished examining or understanding the work" and that some researchers "might still need to assess the mathematical ideas behind the proof".
The rebuild: 175,187 modules, zero errors
The first and most basic check was whether the code compiles at all. A Lean proof is a long chain of tiny logical steps, and Lean's kernel — the small core program that does the final checking — verifies each one against a short list of rules. If a single step fails, the whole thing fails. Sutherland reported that all 175,187 modules in Alpöge's Lean archive compiled without errors, and that the proof had been "independently rechecked by the Lean kernel with only the three standard axioms". Those three axioms — propext, Classical.choice and Quot.sound — are the standard foundations Lean normally relies on. Their presence, and nothing else, means no extra assumption was slipped in to carry the result through a side door.
Sutherland's assessment was blunt: "Honestly the chance that anything is fishy here is approximately zero." That is a strong statement from a careful mathematician, and it reflects how much work the kernel check does. A human referee can miss a subtle gap in a 300-page paper. A kernel cannot be charmed, tired or rushed. It checks everything, or it refuses the proof. The export involved is enormous — reports describe a 61.6 GB exported proof — and the checking required serious computing resources, with one verification page noting that running the full check yourself needs around 100 GB of memory. Dozens of people reported that the library compiles and produces the same exported proof, according to the Lean community statement.
The comparator run, explained
Imagine a student who is asked to prove a hard theorem and instead quietly redefines one of the symbols in the question so the theorem becomes trivial. The proof would be flawless — and worthless. Kevin Buzzard of Imperial College London, one of the best-known advocates of Lean in mathematics, put the danger first when judging any Lean proof that arrives as a file. In a blog post quoted by Scientific American, he wrote: "The first question which one has to ask when judging a PDF which comes with some Lean code is: 'what is the statement of the theorem in the Lean code, and does this correspond to the actual theorem statement.'"
The comparator run is the check built to answer exactly that question. It compares the theorem as stated in a trusted file with the theorem delivered through the untrusted imports — the millions of lines of code that could, in principle, redefine a symbol and prove something easier. It also checks that only the standard axioms are used. Buzzard described what the run would establish: it checks "(a) the code does not do anything nefarious like redefine the meaning of any of the symbols in the statement and (b) (assuming it's run in paranoid mode) the code compiles with several independent kernels (so the code is not just exploiting an as yet unknown bug in Lean's official kernel)". The run typically takes a matter of hours.
What con-leche's compare mode confirmed
The clearest confirmed result on the statement-matching question comes from con-leche, one of the independently written checkers. According to Andrew Sutherland's verification page, as summarised on 21 November, con-leche's `–compare` mode on the full export accepted all five headline theorems as "justified by the solution", exiting with code 0 at 21:19:12 UTC on 20 November. In other words, an independent program compared the statements in Alpöge's submission against the challenge statements and found that they match. A further con-leche run in `–verified` mode on the full export accepted 1,468,418 declarations.
That is the heart of what changed between Friday and Saturday. The proof is not only internally consistent; the statements it proves are the statements mathematicians actually asked for, including the Riemann hypothesis itself and the generalised version for Dirichlet L-functions. The main file proves the Riemann hypothesis exactly as it is already defined in Mathlib, rather than a version written by the authors themselves — a point that matters enormously, because Mathlib's definition has been public since 2023 and was not written to suit this proof. The formal statement sits under a comment in the library noting that constructing a term of that type is worth a million dollars.
What was still running
Here the reporting is less tidy than the headlines suggest, and honesty requires saying so. Scientific American's updated report states plainly: "That run has now been completed." But a detailed tracker of Sutherland's verification page, updated at 14:00 UTC on 21 November, recorded that the end-to-end `lake comparator` run on the main export was still running at the 12:55 UTC refresh, even though its statement comparison and axiom check had already passed once, and even though Lean default and paranoid kernels on the export were also still running. The same tracker noted that con-leche's compare mode — an independent statement-matching path — had already accepted all five theorems.
How should a reader reconcile those two accounts? The most careful reading is that the statement-matching question has been answered in substance by an independent checker, while the single official end-to-end `lake comparator` run was, at the last timestamped update we could verify, still in flight. It may well have completed after that update; different outlets were working from different snapshots of a fast-moving verification effort. What is certain is that the comparator question is no longer open in the way it was on Friday, when nothing had matched the statements at all. What remains uncertain is only the precise timing of one particular run's completion — a bookkeeping detail, not a hole in the proof. We flag the discrepancy in the "What We Could Not Verify" section below rather than smoothing it over.
Three independent kernels, one verdict
Beyond compilation and statement matching, there is a third layer of checking: making sure the result does not depend on a bug in Lean itself. If Lean's official kernel had an unknown flaw, a proof could pass through it while being wrong. The defence is to run the proof through checkers written independently, by different people, in different languages, sharing as little code as possible with Lean.
Three such checkers have now accepted the exported proof. Nanoda, an independent Rust implementation of the logical checker, accepted the full export in 27 minutes and 42 seconds. Con-leche accepted it in 3 minutes and 15 seconds. And lean4lean — a checker written in Lean itself — passed on every compiled module: 175,187 of 175,187, with five modules needing a raised step limit on reruns. Sutherland's refreshed summary line reads that nanoda, con-leche (including `–compare` against the challenge) and lean4lean all accept the published artifact.
What the proof actually says — and how big it is
The Riemann hypothesis, first proposed by Bernhard Riemann in 1859, concerns the distribution of prime numbers — the numbers divisible only by themselves and 1. Riemann linked the pattern of the primes to the points where a function now called the Riemann zeta function equals zero. The hypothesis says that every non-trivial zero of that function has the same real coordinate, exactly one-half. If true, the count of primes stays about as close to a smooth curve as it possibly can, and hundreds of papers already prove results on the condition that it holds. Computers have checked the first ten trillion of those zeros, and all of them sit on the line — but there are infinitely many, so checking can never settle the question.
Alpöge's claim goes beyond the original hypothesis. He says the same method also settles the Generalised Riemann Hypothesis for Dirichlet L-functions, a broader family of related functions. Sutherland's verification page narrows that claim in one important respect: the "GRH" in the code covers Dirichlet L-functions only, and makes no claim about more general L-functions such as the Dedekind zeta function. Readers should not overstate the scope. The achievement is vast as stated; it does not need inflating.
The sheer scale of the formalisation is itself part of the story. The Lean archive contains about 17,374,215 lines of code in 175,187 files, split into 14 parts. Alpöge has said that a conventional written version of the argument could run to more than 250,000 pages. He also writes, by his own account, that he has not yet mastered all of the mathematics around it and intends to write it up properly when he is able to simplify it. When he tried an AI-written version of the full write-up, he found himself staring at what he described as "some massive, unreadable travesty of 'slop'". That candour is striking, and it frames the central oddity of this moment: the mathematical community now has a proof it believes is correct, in a form almost nobody can read.
What machine verification does — and does not — prove
It is worth pausing on what has actually been established, because the headlines have sometimes run ahead of the substance. Four things are now solid. First, the Lean code compiles, in full, without errors, on machines outside Anthropic. Second, the statements proved are the right statements, as confirmed by an independent comparison. Third, only the three standard axioms are used. Fourth, three independently written checkers accept the proof. That is an extraordinary amount of evidence, gathered in about a day, and it justifies the Lean and Mathlib communities' confidence.
What has not been established is equally important. Nobody has yet read the argument and understood it. There is no human-readable paper, no peer review, and no journal publication. The Clay Mathematics Institute still lists the problem as unsolved, and its one-million-dollar prize process has its own rules, which a Lean formalisation alone does not satisfy. The history of this problem counsels humility: claimed proofs have a long record of collapsing, including Louis de Branges de Bourcia's 124-page proof announced in 2004 and Michael Atiyah's attempt in 2018, neither of which convinced specialists. This case differs from those in one decisive respect — it comes with machine-checkable code that has survived independent verification — but the difference is in the evidence, not in the need for human understanding. OpenAI reported a similarly machine-verified proof of the Yang-Mills mass gap earlier in the week: OpenAI Yang-Mills Proof Verified.
There is also a philosophical objection that verification does not answer. Maksym Radziwiłł, a mathematician quoted by Quanta Magazine, put it this way: "The object of mathematics as the rest of science is really the question 'why?' and this doesn't answer it at all." A 17-million-line formal proof tells us that the Riemann hypothesis follows from the axioms. It does not, at least not yet, tell us why the primes behave this way. For many mathematicians, that "why" is the point of the subject. The proof may be correct and still leave the deepest question untouched — for now.
The rival claim: a human proof, under revision
The weekend produced a second drama that should not be lost in the excitement over Alpöge. On the same day, Ravi Vakil, president of the American Mathematical Society and a mathematician at Stanford University, posted a draft by Alessandro Masullo, a young Stanford researcher, claiming the Generalised Riemann Hypothesis by an entirely human route, with no AI assistance. According to Vakil, the proof had been in the hands of Brian Conrad, a Stanford mathematician and an expert on the subject, who had been giving Masullo feedback over several months. Conrad confirmed to Scientific American that the details of Vakil's blog post were true.
The two claims are now at very different stages. Alpöge's has cleared several important computer-based verification checks. Masullo's has encountered questions from mathematicians examining the draft. Asked directly whether he thinks Masullo's work will prove valid, Conrad said more work is needed: "To judge the success of the method requires the paper to reach a stage of more final revision. It's undergoing major revision." He also explained the choice to post it now as a decision between "two unpalatable options": announce it, or "remain silent and let a period of time go by, only to post something later."
Reaction: wonder, caution and awkward timing
The reaction has ranged from awe to sharp scepticism, sometimes from the same people. Elon Musk was reported by BigGo Finance to have called the development "mind-blowing", a reaction that captures the public mood even as specialists urge restraint. James Maynard, a Fields Medal winner at the University of Oxford, told Nature: "If this fully checks out, it is clearly a historic milestone in mathematics." Hector Pasten of the Pontifical Catholic University of Chile told Scientific American that, if the proof is correct and its ideas are understood, it could connect different areas of mathematics that researchers have only partially linked.
The caution has been just as prominent. Andrew Granville, a number theorist at the University of Montreal, told Nature that Alpöge should have taken the time to prepare a clearer write-up before making the announcement, calling what was released "garbled". Granville also raised the timing: the announcement came just one day after Anthropic's shares began trading on Nasdaq, and days after the company's initial public offering, which initially valued the company at US$1.5 trillion before its valuation dropped on 19 November. "There is an obvious cynical view, to do with the timing, but it is false," Alpöge wrote in his note. Anthropic itself declined to comment on Friday as experts rushed to check the work, according to The Wall Street Journal, which also reported that the Lean and Mathlib communities had finished verifying 17 million lines of code and issued their ruling on Saturday.
What still has to happen
First, the remaining verification runs need to finish and be reported cleanly. The independent checkers have accepted the proof, but the full end-to-end `lake comparator` run and the Lean kernel replays on the export were still in progress at the last timestamped update we could verify. Those should close out in the coming days, and any discrepancy between outlets about their status should be resolved by the primary source — Sutherland's verification page — rather than by headline.
Second, mathematicians need to understand the argument. Tao's warning that digestion "will take some time" is probably an understatement. Alpöge himself says he has not yet mastered all of the mathematics around the proof. The work of understanding will likely take months or years, and it will involve many mathematicians, not one. The most valuable output of that process may not be a verdict — the verdict already looks settled — but the new ideas the proof contains, and whether they can be shortened, generalised or applied elsewhere.
Third, a human-readable account needs to be written and peer-reviewed. Alpöge has said he plans to write the proof up properly when he is able to simplify it. Until that exists, the proof will remain, in a real sense, unread — a strange status for the solution to the most famous problem in mathematics.
None of that diminishes what happened on Saturday. A claim that arrived on Friday as a four-page note and 17 million lines of code is now certified correct in its formal form. The statement is right, the proof checks out, and the chance of a trick is, in Sutherland's words, approximately zero. What remains is the harder, slower and ultimately more important work: understanding it.
What We Could Not Verify
- Whether the full end-to-end `lake comparator` run had finished. Scientific American reported that the comparator run "has now been completed", but a detailed tracker of Andrew Sutherland's verification page, updated at 14:00 UTC on 21 November, recorded the `lake comparator` end-to-end run on the main export as still running at the 12:55 UTC refresh, with its statement comparison and axiom check already passed. We could not reconcile the two accounts from a single primary source at the time of writing, so we report both and treat the statement-matching question as settled in substance by con-leche's `–compare` result, with the timing of the official run's completion unresolved.
- Elon Musk's exact words. BigGo Finance reported that Musk called the development "mind-blowing". We could not verify the exact wording or the original post from a primary source before publication, so we attribute the description to that report rather than presenting it as a direct quotation.
- Which Claude model was used. Alpöge's note names no model version, and Anthropic had made no official statement at the time of reporting. We do not state a model version anywhere in this article.
- The full contents of the Lean and Mathlib joint statement. We quote the two sentences carried by Scientific American and USA Today. We could not verify the complete text of the statement from the Lean community blog directly before publication.
- The status of the Masullo draft beyond Conrad's remarks. We report that mathematicians have raised problems and that the draft is undergoing major revision, as stated by Brian Conrad. We could not verify the specific mathematical objections, which have not been published in a single citable source.
- Any timetable for a human-readable proof or Clay Institute recognition. No source we could verify gives a date for either. We do not offer one.