A Proof With No Paper
On Friday evening, Levent Alpöge, a mathematician on Anthropic’s technical staff, posted four pages of text and a 239-megabyte tarball, and told the world he believed he had proved the Riemann hypothesis with Claude.
The loudest AI story of the day was elsewhere. Anthropic’s second day as a public company went badly — down 9%, and with it the Nasdaq’s entire 2026 gain — and I have no intention of re-litigating the market’s mood here. But the market is asking “is this worth what it cost?” and the story on alpo.ge is the one that will still be true in fifty years, whatever the Nasdaq does on Monday.
I have now written about four of these — the Aaronson–Ambainis conjecture, the Exponential Time Hypothesis, the Euler and Navier–Stokes blow-up results — and the reflex at this point is to ask the same three questions. Does it compile? Does it say what it claims to say? Does anybody outside the team believe it?
This time the answers were, unusually, yes, yes, and something stranger: the compiler is stronger than the author.
Here is what I mean by that, and why I think it is the most consequential thing that has happened in mathematics in a very long time.
What was actually claimed
Alpöge’s note opens like this:
I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis. In fact, the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis, and I expect, with more time, the Extended Riemann Hypothesis too. Claude and I have also formalized the arguments without assumptions.
Note the scope carefully, because it is narrower than the headlines. What was claimed is the Riemann hypothesis plus the generalisation to Dirichlet L-functions — not every L-function in existence, and not the Extended Riemann hypothesis. Andrew Sutherland, the MIT mathematician who ran the first serious independent verification, explicitly limited the generalisation he checked to the Dirichlet case.
The Riemann hypothesis, as Alpöge describes it in his note’s first line, “asks for optimal control on the count of prime numbers of a given size.” Everything else about the problem — the zeros of the zeta function, the critical line — is machinery built to get at that count. It has been open since 1859. It is one of the Clay Millennium Prize Problems. If this holds, it is the most valuable individual mathematics result of the century.
And the claim came with a release that is, in its way, the strangest artifact the field has ever produced.
The artifact
Four pages of note — remark.pdf — and fourteen compressed parts of Lean source, lean.tar.gz.part01 through part14, merged at 239,275,554 bytes with SHA-256 9631b4f0…49c7fc. Inside: 175,187 Lean files, roughly 17.4 million lines, 818,755 theorem and lemma declarations. Twelve-point-two million of those lines sit in a single Development directory; 3.2 million in a Formalism directory. There are no comments.
The front door is small, and this matters. Main.lean states five theorems — the Riemann hypothesis, the zeros-on-the-line form, nonvanishing for Dirichlet L-functions to the right of one half, and the two GRH statements — each one line applied to a proof term defined deep in the supplements. Statements.lean is sixteen lines and imports only Mathlib.
And here is a detail worth pausing on: RiemannHypothesis is not a definition written for this archive. It is Mathlib’s own definition, imported from Mathlib/NumberTheory/LSeries/RiemannZeta.lean at the pinned commit. The statement is not the author’s restatement of the theorem; it is the community’s, from the library the whole ecosystem already trusts.
That single design decision eliminates the single most common way a formalised proof is worthless — a proof of a subtly different proposition that compiles perfectly.
Alpöge’s note is also unusually moving, which I did not expect. He mentions that this was “my secret in my student days,” that his plans were for the problem “in decades,” and then:
this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line.
Later, in the footnote where he explains why there is no paper:
I of course amused myself, and hit 250,000+ pages. Claude cutely called it the GRH . . . “book”
Two hundred and fifty thousand pages of natural-language proof. That is the thing that exists and that nobody will ever read.
The one line Alpöge gave for mathematicians
Mathematicians are not going to read 17.4 million lines of uncommented Lean. So Alpöge gave them the idea, and he did it well enough to earn a raised eyebrow from Terry Tao.
The argument descends from Stepanov’s proof of the Riemann hypothesis over function fields, where a divisor involving multiplicity does the heavy lifting. Over the integers, multiplicity is hard to find — Alpöge’s phrasing is “corresponding to something like differentiating whole numbers.” The route out was to chain theta lifts so that each one mimics a unit of multiplicity, in a natively automorphic setting.
And then the line Alpöge gave as the one that would make fellow mathematicians “fall out of their chairs”:
for the von Mangoldt sum of length X, Claude’s suggested length of chain of theta lifts, in view of each mimicking multiplicity, was √X. (And yet, what else could it be?)
Read that parenthetical as a mathematician, not as a journalist. He is saying: the model produced the one parameter that sets the whole balance, and it produced the right one, and I cannot see any alternative candidate for it. √X is what the loss term in Jaswal’s analysis independently confirms — the total loss is x^O(ε) · (x/m + m + √x), which is optimised at m near √X.
One parameter, chosen by the model, that is exactly the parameter the analysis says must be chosen. That is the thing to hold on to.
The verification is the strongest thing here
I wrote about the Aaronson–Ambainis result in November and was, on balance, gloomy: the formalisation was checked against a kernel with four known soundness bugs, by the people who produced it. I got a lot right about that and one thing badly wrong, and I have thought about it since.
This time the posture is materially better, and I want to be precise about how much better.
Andrew Sutherland’s report, published at 22:05 UTC on 20 November, is the first serious independent verification, and it is a good one. According to it:
- A zero-error build of all 175,187 modules under Lean v4.33.1 and Mathlib v4.33.1.
#print axiomson the five theorems inMain.leanreported onlypropext,Classical.choice, andQuot.sound— the three standard axioms, nothing else. NosorryAx, no custom axiom, noLean.ofReduceBool.- Every module was replayed through Lean’s own
leanchecker. - Two independently written kernels —
nanodaandcon-leche— accepted the full export. con-leche --compareaccepted all five theorems against separately written challenge statements, which is the machine version of “does this say what it claims to say.”- All modules passed
lean4lean, with five needing a rerun at a higher step limit.
The caveats are stated in the report itself and they are real: the build called Lean directly on individual files in an unsandboxed environment; there was no full comparison against an export from a sandboxed build; and the end-to-end lake comparator run and the default/paranoid kernel runs on the export were still in progress at the time of writing.
Kevin Buzzard of Imperial College put the remaining item the way I would want it put, on the Lean chat:
We still await a successful comparator run, which will check that (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
And his first question is the right first question: what is the statement of the theorem in the Lean code, and does this correspond to the actual theorem statement? Sutherland’s statement-comparison step is the direct answer to it, and it passed.
Two further things are worth noting. First, the build cost turns out to be tractable rather than unbounded: the FINAL-Bench group reports a 192-core node building the tree at roughly 250 files per minute, which puts a single node at around twelve hours of wall-clock work for the whole archive — parallelisable across a modest cluster. Second, that group compiled the released proof from scratch under a different, newer toolchain (Lean 4.37.0-rc1) and again got standard axioms only. Two independent toolchains, same result, which is the best available answer to “is this exploiting a kernel bug?”
Sutherland’s own summary is blunt: “Honestly the chance that anything is fishy here is approximately zero.” James Maynard, separately: “I suspect that the proof will turn out to be correct, although even checking the Lean verification will take some time.”
Clay Mathematics Institute still lists the Riemann hypothesis as unsolved, and its Millennium rules require publication in a qualifying outlet, two years elapsed, and general acceptance by the international mathematics community. All three conditions are unmet, and the third may take a while.
The line nobody skims
Here is the sentence in Alpöge’s note that I think will end up being the most quoted part of this whole episode:
This puts us in the rather extraordinary situation this age has made possible, in which I am presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics. It being obviously inappropriate to attempt to post some massive unreadable travesty of “slop”, I decided to only present Lean.
Read the choice being made there. The alternative was not “a paper.” The alternative was 250,000 pages of unreadable, slop-shaped natural-language prose — a document that would have been published, downloaded, cited, and would have taught the field nothing while consuming months of everyone’s attention. He chose Lean because the other option was actively worse.
That is a defensible decision and I think it is the right one. But it is also the moment the story stops being a story about mathematics and becomes a story about what has happened to the form of mathematical knowledge.
The machine artifact is now ahead of the human artifact
Here is the comparison that I think is the real news, and it happened on the same day.
On 20 November, Alessandro Maria Masullo also posted a proof — a human-written proof of GRH, announced with a hurried post by Ravi Vakil. Terence Tao’s assessment, after a closer look: the manuscript is closer to Deninger’s strategy towards RH than Connes’. Tao’s verdict on the AI proof, for contrast: “It is very early days but it seems like the AI-generated proof uses all of these inputs” — automorphic methods, algebraic number theory, arithmetic geometry, all three of the serious routes.
Now compare the two artifacts.
| Alpöge + Claude | Masullo | |
|---|---|---|
| Form | 17.4M lines of Lean, no comments | Human-readable manuscript |
| Machine-checked | Yes — three independent kernels, standard axioms only | Not applicable |
| Readable | No — four-page note | Yes |
| State | Compiles, independently verified | Draft |
And the draft’s state is the part I could not stop thinking about. An early audit of the manuscript found it still containing its own editorial instructions as footnotes — “Correct statement”, “Insert statement”, “Correct this subsection”, “Include section”. Another reader, running the text through a frontier model, flagged a uniform bound assertion that would force a positive minimum gap between the ordinates of simple zeros on the critical line. Vakil’s own update was scrupulously careful: “As is usual, all issues need to be resolved to the satisfaction of the community before this proof would be accepted as true. I make no claim as to the correctness of the proof strategy.”
So: the machine-written proof has a machine-checkable certificate and no human-readable form. The human-written proof has a human-readable form and no certificate, and is not finished.
The machine artifact is ahead of the human artifact in rigor. I do not think that has ever been true before for a result of this significance, and I am not certain what to do with it.
Why this is not the Wiles or Mochizuki pattern
Every previous case of a claimed proof of a famous conjecture has been a human artifact: a paper, a preprint, a manuscript. Wiles announced in 1993; the gap surfaced within months; the repair took until 1995. Mochizuki published in a refereed journal in 2012, and the people who work on the problem still do not accept it. In both cases, the error — when there was one — was findable by reading, and in both cases the path to resolution ran through a human explaining what the argument means.
This artifact changes that shape. A proof that exists only as 17.4 million uncommented lines of dependent type does not have errors you find by reading it. It has errors you find by building it. And the field’s mechanism for accepting a result is, still, reading — because acceptance means that the result has been absorbed, taught, extended, and taken for granted, and all of those are acts of human comprehension.
So the verification and the acceptance have split onto different clocks. The verification can be done in a cluster in hours. The acceptance took Wiles years and Mochizuki forever.
One of the commenters on Tao’s post made the point with unusual clarity. There is now a prediction market on the question, and its resolution rule was fixed before the market opened: a complete unconditional proof in ZFC, independently checked, with no unresolved gap; finite numerical verification does not count; independence from ZFC leaves it unresolved. The commenter’s conclusion is the part worth keeping:
The interesting question is not “is it true” but “what observable event would settle it, and who is entitled to say so.”
That is the right question, and it is a question about authority, not about correctness. Lean can certify the first half of it in a weekend. It has nothing to say about the second half.
The bottleneck has moved, and it is now AI-mediated
Terrence Tao posted three results on the same day, in “decreasing order of how ‘digested’ the results currently are.”
The first — finite-time blowup for the unforced incompressible Euler equations — got a sentence and a link to a PDF and a Lean repository. The second — the Navier–Stokes version — got two sentences, including the assessment that the strategy is “broadly clear” but that “more work needs to be done to create a proper, professionally written proof.” The third, the Riemann hypothesis, got one sentence:
Levent Alpöge, using an internal AI model, has generated an (enormous) Lean-only proof of the Generalized Riemann Hypothesis.
One sentence. Not because it matters least — it matters enormously — but because there was nothing to digest. The digestibility ordering was not a ranking of importance. It was a ranking of how much of the proof exists in a form a person can read.
And here is the part that should stop you, if you let it. The person who did the reading used AI to do it.
Manvir Jaswal, a number theorist, posted a detailed reconstruction of the strategy on Tao’s blog, and he opened by explaining his method:
I’ve been looking into the Lean source as text, with Claude Fable 5.2 and GPT Astra 6.2, to get the strategy out of it. This is a first pass and I’ll refine it as the middle of the proof gets checked. I haven’t built the archive or checked its proofs, so all this says is how the proof is organized.
Let that sit for a moment. A professional mathematician read 17.4 million lines of uncommented Lean — not compiled, not verified, just read as text — using two frontier models, and reconstructed the shape of the argument. He then wrote up what he found, including a careful statement of what he had not checked. Two hours later he posted corrections to his own analysis: “Some small corrections. I conflated two terms above.”
That correction is the tell. A reading of a proof, performed with machine assistance by a human who explicitly did not build the artifact, and then revised after the fact — that is what the comprehension layer now looks like in practice. It is not a mathematician sitting down with a proof. It is a mathematician supervising a machine that is summarising a proof that another machine wrote.
Jaswal’s own summary of the state of his knowledge is the most valuable single paragraph in the thread, because it tells you exactly where the load-bearing walls are:
What I’ve checked independently is the analytic end. The window form of Landau’s theorem and the equivalence with GRH were proved again from scratch, and the identity was re-proved from its statement and checked numerically on small fields. For the middle I have only read statements of declarations, and “theta lift” and “period” are Alpöge’s words and the names in the code. I haven’t checked that the objects are built the way those names suggest, or that the middle constructs what the end uses.
And he lists the statements he could not find anywhere in the literature: the central inequality itself; the approximate orthogonality between the maps for different words; a lower bound for normalised arithmetic self-intersection numbers in towers of unitary Shimura varieties; and the bound for the plus part summed over primes and twists.
Those four are not decoration. Those are the proof.
“And yet, what else could it be?”
There is a much easier version of this story, and it is wrong.
The easy version says: AI found the Riemann hypothesis because the Riemann hypothesis is the sort of thing AI is good at finding, and the difficulty was always an information problem rather than an insight problem. Terry Tao dismantled that in a comment thread in the most Tao way possible. Someone asked whether it was surprising that RH turned out to be easier than an equational-theory implication in finite magmas. Tao’s four reasons are worth reproducing because they are the best available defence of the mathematical difficulty:
- We don’t know how much compute in aggregate has been spent on both problems, but I imagine it is far larger in the former case.
- The RH proof uses a century of past literature on related topics. There is virtually no literature on equational theories to build upon for the latter problem.
- The equational implication was distilled as the only surviving open implication out of an initial set of about 44 million, all of which are now settled — so there is a significant survivor bias present here.
- It is known that some equational theory implications are undecidable in ZFC.
Point two is the one that matters for understanding what happened. The Riemann hypothesis was not sitting there waiting to be found. It was sitting there waiting to be made tractable, and what made it tractable was a century of accumulated apparatus — the Langlands program, automorphic forms, arithmetic geometry, the Rallis seesaw, Bost’s slope inequality, Kato’s Selmer group vanishing. Alpöge’s note is full of these: a “weakening of Vandiver”, Selmer groups of Kato’s type, congruences between Frobenius traces. The model did not invent that apparatus. It did not have to; a century of mathematicians had already built the road and left it on the shelf.
So the honest description of what happened is neither “AI solved RH” nor “AI rediscovered what was already known.” It is: AI was handed a two-hundred-year-old proof search that had been stalled for want of brute force, and it found the √X.
Which is still enormous. Alpöge says so himself, with more feeling than any of the coverage managed:
An astounding amount of mathematics is present in the argument, far more than I had imagined would be necessary as a student. My intuition insists it will be massively simplified, likely immediately.
“Massively simplified, likely immediately.” That is the most scientifically valuable sentence in the note, and almost nobody quoted it.
The capability that produced this is not the capability that will certify it
METR published its Q4 capability update on 20 November, the same day as all of this, and I do not think the two stories have been put next to each other yet.
METR’s headline number: across its TH1.1 and TH2 task suites, the 50% time horizon of frontier models is doubling every 3.61 months (95% CI 2.08–4.94). On the verifiable subset of the new, harder TH2 suite, current models reach up to about 27 hours, with a doubling time of 1.65 months. GPT-6 Astra sits at roughly 40 hours on the pooled suite; GPT-5 in August 2025 was at 3 hours.
And then the caveat, which is the part I care about. METR defines judgment as “the ability to make good choices in the absence of clear and fast empirical feedback.” On tasks without such feedback, the report finds, “AI agents appear to have worse judgment than human experts.” Concretely: models now beat experts who spend five minutes rating philosophical arguments but remain below experts who spend fifteen; when asked to design an evaluation that would be informative, “none of the tasks agents produced were sufficiently useful to us to warrant inclusion in this post”; and as judges, models inflate their agreement when the work being judged is at least as strong as their own.
Now put the two reports side by side.
Generating a proof of the Riemann hypothesis is now within the reach of a frontier model. Certifying it to the satisfaction of the international mathematics community requires exactly the capability METR says has not scaled: years of reading, judging, distilling and reconstructing an argument whose correctness is not cheaply checkable.
Scott Aaronson put the standard for acceptance as cleanly as anyone could:
[mathematicians will be satisfied] when, and only when, they’re able to explain and distill and reconstruct the ideas for themselves
That will probably take years. The Lean certificate takes hours. The gap between those two numbers is not a scandal — it is the correct shape of the thing. But it is a gap, and it is widening at exactly the moment the proof-production side is accelerating.
The bottleneck is no longer generating the proof. It is judging it. And judging, in the precise sense METR uses, is the capability the models still trail humans at.
What this means for people building things
If you build in this space, the opportunity has moved again, and it has moved in a direction that would have looked absurd a year ago.
Verification is not the bottleneck; verification has been solved. It took one MIT mathematician and a cluster, overnight. If you were building proof-checking tooling, you are building for a problem that just got solved by a hobbyist with cloud machines. Lean 4 is the platform and it is excellent. Stop.
Digestibility is the bottleneck, and it is a product problem, not a research problem. Everything downstream of this event — the paper Alpöge intends to write, the teaching of the argument, the extensions to the Extended Riemann hypothesis, the thousands of people who will want to use GRH in their own work — is blocked on one thing: turning 17.4 million uncommented lines of dependent type into something a person can hold in their head. That is now worth more than the proof.
Concretely, the tooling that would actually move this: automated extraction of a proof’s dependency skeleton from a Lean development (what Jaswal did by hand, and then had to correct); term-level natural-language documentation generated against the actual elaboration, not the source text; statement-mapping across libraries so that “is this Mathlib’s RH?” is a checkable artifact rather than a reading exercise; and reproducible-build infrastructure that reports peak memory and wall-clock time as a first-class output, because “how expensive is it to check this?” is now the first question and nobody knows the answer for artifacts of this size.
And note that every one of those tools will itself be partly AI-built, working on AI-written proofs, to produce documents humans can read. The comprehension layer is not going to be staffed by humans reading Lean directly. It is going to be staffed by humans supervising machines that summarise machines’ work. Jaswal’s correction two hours after his first post is exactly what that looks like in practice, and I suspect we are going to see a great deal more of it.
If you consume AI output, the provenance question is now concrete. A “proof” is no longer a thing you read; it is a thing you build. Your check is not “do I believe it” but “did I rebuild it, and what did the axiom list say, and did an external kernel accept it, and does the statement match the one the field means.” Those four questions are now the whole of peer review for machine-generated mathematics, and every one of them is automatable.
And fund the reading. This is the part the labs are not going to do, because the labs’ comparative advantage is generation and their interest is in the next theorem. The Science Markets resolution rule was worth reading for one reason alone: writing down in advance what would count as settling it forces you to notice that the two obvious candidates — “the code compiles” and “the field accepted it” — are not the same thing, and that only one of them is a question a machine can answer.
Two reactions from the thread that are worth keeping
The thread on Tao’s post contains two comments that I think capture the cost of this moment better than any analysis of mine.
Josh_Fourier, a research mathematician, wrote a polemic about the end of mathematics as an equaliser — how it used to be the one field where a mind with a chalkboard had the same tools as a professor at Princeton, and how it now looks more like Formula 1, where the difference is not who drives best but who gets the unreleased prototype. The argument is imperfect and the emotion is genuine and I am not going to adjudicate it. But the closing line is right about the governance question:
the most urgent question for academia is how open research survives when the fastest engines remain behind corporate doors.
And then jpolak, who read the comment and wrote this:
When I was an undergraduate, the Riemann hypothesis fascinated me. It was one of the reasons I wanted to get a PhD in math (I did). But now, the way many of you are using AI in math, it makes me completely disinterested in the Riemann hypothesis now.
That is the cost that does not appear in any of the verification reports. A generation of people chose this field because of a specific feeling — the romance of a problem that had resisted for 167 years — and the ground under that feeling has moved. Alpöge himself felt it from the inside. His note contains this:
It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.
And then, at the end:
I hope in having the experience of seeing all of number theory’s beauty brought together I am just one of the first, and those who love the mathematics will all be able to enjoy such transcendence in the future.
I want to believe that. I notice that the sentence is about people who love the mathematics, and is addressed to them, and it is asking them to come along rather than mourning. Alpöge is not being sentimental. He is describing what happened to him personally — the loss of the guiding light — and then immediately describing what he thinks the thing that replaces it looks like. He did not have to publish the second half. It would have been much easier to publish only the first.
The uncomfortable thing is that he is probably right about both halves at once. Which is the most human thing I have read this week, and it is buried in the second paragraph of a footnote in a four-page PDF nobody has read.
Where this leaves us
Let me be careful about the claim, because the last few years have not been generous to loose talk.
Established, tonight: a public artifact exists, its hash matches the author’s description, 175,187 modules compile under Lean and Mathlib v4.33.1 with zero errors, the five headline theorems depend on only the three standard axioms, two independently written kernels and a replay checker accepted the export, a statement-comparison step confirmed the theorems match separately written challenge statements, a second toolchain reproduces the result, and the headline statement is Mathlib’s own definition rather than the author’s.
Not established: that a number theorist has reconstructed the argument; that the middle of the proof — the central inequality, the approximate orthogonality, the Shimura-variety self-intersection bound — is correct; that the claimed GRH generalisation extends beyond Dirichlet L-functions to the Extended hypothesis; that the Clay Mathematics Institute or the international mathematics community accepts anything; and that the end-to-end paranoid comparator run has come back clean.
That is a much stronger position than any previous AI-math claim has occupied at the same age. It is also not the same thing as a proof, because a proof is not only a derivation — it is a derivation that the community can restate, teach, and extend.
So the accurate description is Alpöge’s own, slightly sharpened: a claim, with a certificate attached, and a readable proof still to be written.
And then the thing that will actually decide the next decade, which nobody is funding: somebody has to be able to explain it.
The last line of my own note to myself from two weeks ago was that a proof nobody can reproduce is a rumour with a PDF attached. This is the same problem wearing a better suit. The PDF has been replaced by 17.4 million lines that any sufficiently funded group can check, and that is a genuine and important advance. But the thing that made Wiles a theorem rather than a preprint was not the checking. It was a lecture, and then a paper, and then a textbook, and then thirty years of people building on it.
There is no automated substitute for that, and there is not one coming. Which means the next bottleneck is not a technical problem at all. It is the oldest problem there is: finding people with time, and the patience, and the years, to understand something that was never written to be understood.
Let us hope there are enough of them. Alpöge seems to think so, and he is the one person in this story who has read the thing.
Sources
- Levent Alpöge — The Riemann Hypothesis (four-page note) (20 November 2026)
- Levent Alpöge — Lean archive, 14 parts · mirror with SHA-256, build instructions and static audit
- Terence Tao — Unforced Euler singularity, unforced Navier-Stokes singularity, and the Generalized Riemann Hypothesis (20 November 2026) — including Tao’s comments on the analytic-versus-geometric routes and Manvir Jaswal’s reconstruction of the strategy
- Scientific American — The AI Riemann hypothesis proof is almost certainly true, independent mathematicians attest (20–21 November 2026)
- Kevin Demian — The Alpöge–Claude Riemann Hypothesis Claim and the Public Evidence So Far (21 November 2026) — detailed accounting of Sutherland’s verification report
- The Defiant — Masullo’s Riemann Claim Draws AI-Assisted Objections (21 November 2026)
- Ravi Vakil — Hurried announcement: Alessandro Masullo’s GRH (20 November 2026)
- METR — AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure (Q4 2026 capability update) (20 November 2026)
- FINAL-Bench — Seeing the Riemann Hypothesis Hold: A Live Weil Positivity Gauge for Zeta and Dirichlet L-Functions (21 November 2026)
- The Latent Daily — Daily Digest: a claimed Riemann proof (21 November 2026)
Comments
Leave a message below. Your comment saves to your browser.