KKDEMIAN

The Alpöge–Claude Riemann Hypothesis Claim and the Public Evidence So Far

Alpöge released a four-page announcement and 14 Lean archive parts. This report distinguishes RH from Dirichlet GRH, reviews the independent machine checks and explains how Clay considers a Millennium Prize solution.

Research

On November 20, 2026, number theorist Levent Alpöge announced that he believed he and Claude had found a proof of the Riemann Hypothesis. He said that the same method resolves the generalized hypothesis for Dirichlet L-functions. He published a four-page note and the Lean source in 14 archive parts. By 22:05 UTC that day, MIT mathematician Andrew Sutherland had reported successful compilation, an axiom audit and independent kernel checks. His end-to-end Comparator run was still in progress, and the Clay Mathematics Institute still listed the problem as unsolved. This article checked the public record around 01:20 UTC on November 21, 2026. The machine results below retain the timestamp of Sutherland’s report. Alpöge’s note · Sutherland’s verification log · Clay’s problem page

What the hypothesis says

For complex s with real part greater than 1, the Riemann zeta function is defined by Σ(n ≥ 1) n^(−s) and can also be expressed as an Euler product over the primes. After continuation to the complex plane, it has trivial zeros at the negative even integers −2, −4, … and a pole at s = 1. The hypothesis, proposed in 1859, places every nontrivial zero on the line of real part 1/2. When reading a formal statement, check how it handles the pole and the negative even integers, then confirm that it covers every nontrivial zero. Bombieri’s official problem description

The prime number theorem gives the average behavior of the prime count. Riemann’s explicit formula connects fluctuations around that average with zeros of the zeta function. Under RH, the real parts of the zeros limit their powers of x to x^(1/2). Summing the contributions with the relevant truncation gives error bounds with logarithmic factors. The bound is about the error in counting primes. Finding individual primes still requires separate computation. Dirichlet GRH makes the analogous assertion for L-functions attached to Dirichlet characters and has applications to primes in arithmetic progressions. Bombieri’s official problem description

From a toy model to the announcement

Alpöge’s CV records a Harvard undergraduate mathematics degree, a Cambridge MASt and a Princeton mathematics doctorate in 2020. His dissertation, Points on Curves, was supervised by Manjul Bhargava. He received the Morgan Prize and was a Harvard Society of Fellows Junior Fellow from 2021 to 2025. His earlier work includes integral points on elliptic curves, nonmonogenic number fields and low-lying zeros of Maass-form L-functions. He describes returning to computing and joining Anthropic on his personal website. That background helps explain his choice of problem. The proof still has to be checked on its own merits. Author’s CV · Author’s homepage

Alpöge says he first brought two ideas together in a toy setting and looked for a genuine mathematical setting with the same structure. He recalls Claude moving much faster than he did toward the end. His concrete example is the proposal to imitate multiplicity in Stepanov’s method by chaining theta lifts. Claude suggested a chain on the order of √X for a von Mangoldt sum of length X. Alpöge also says the higher-rank automorphic portions extend beyond what he had mastered. Claude compiled the timeline, and he cautions that his account of some details may be wrong. The account traces how the argument took shape. Its individual steps can be checked against the Lean dependencies and a readable derivation. Alpöge’s note, pp. 1, 3 and 4

Date in 2026 Milestone reported in the note
October 6 Stepanov determinant, the Iwasawa-tower analogy and Bost’s slope inequality
October 20 Shift to a weighted sum built around the von Mangoldt function
November 15 Iterated theta lifts between definite unitary groups, with work on norms, divisibility, the Rallis seesaw and the cost per lift
November 16 The published rough timeline records GRH
November 20 Public release of the four-page note and the formal source

The table rearranges selected milestones by date. The 20 entries in the original are partly out of chronological order. Alpöge says in a footnote that he generated more than 250,000 pages of expanded prose. The figure describes that draft. The public materials are the four-page note and the source archive. The author’s archive runs through part 14. The GitHub link supplied by the reader is a mirror, with a different role from a mathematical check or a license grant. Alpöge’s note

What independent checks established

Sutherland’s 22:05 UTC report separates compilation, kernel checks and statement comparisons. It reports a zero-error build of 175,187 modules under Lean and Mathlib v4.33.1. For the five theorems in Main.lean, #print axioms reported only propext, Classical.choice and Quot.sound. Lean’s leanchecker replayed all modules. The independently written nanoda and con-leche kernels accepted the full export, and con-leche’s --compare mode accepted all five theorems against separately written challenge statements. All modules passed lean4lean, including five that needed a rerun with a higher step limit. Sutherland’s verification files and logs

The same report lists remaining checks and their scope. The build called Lean directly on individual files in an unsandboxed environment. No full comparison with an export from a sandboxed build had been made. The end-to-end lake comparator process and the default and paranoid Lean kernel runs on the export were still under way. Independent machine checks are on record. Further work can narrow what the verification has to trust, confirm how the formal targets correspond to customary mathematical definitions and make the argument usable by other mathematicians. Terence Tao points readers to the Lean community discussion and treats understanding the proof strategy as work still to be done. Sutherland’s verification log · Terence Tao’s first-person account

The scope of GRH checked here

Sutherland limits the reported GRH claims to Dirichlet L-functions. The ordinary RH target uses Mathlib’s own RiemannHypothesis. The other targets cover the zeta-zero statement written out, a nonvanishing result to the right of the critical line for Dirichlet L-functions, and two versions of the critical-strip zero statement. Alpöge describes the broader Extended Riemann Hypothesis as a direction he expects the method to reach with more work. The listed independent checks do not address that extension. Sutherland’s scope of verification · Alpöge’s note

Anthropic published a different result in August. It reported that Claude increased a lower bound on the proportion of zeta zeros on the critical line from 41.6% to 67.2%. A proportion bound and a statement about all zeros have different evidentiary burdens. Even a proportion tending to 100% leaves room for a sparse set of off-line zeros. The November claim concerns the real part of every nontrivial zero. Anthropic’s August report · Clay’s statement of the problem

How Clay considers a solution

The Clay Mathematics Institute still lists the Riemann Hypothesis as unsolved. Its Millennium Prize rules require publication in a qualifying outlet, at least two years since that publication and general acceptance by the international mathematics community before the institute will consider a proposed solution. The date of a personal website release or mirror is insufficient to establish when the two-year clock begins. Formal evidence and a professional mathematical argument can precede the institute’s prize process, which starts under its own rules. Clay’s problem page · Millennium Prize rules

If the remaining checks continue to pass

The immediate number-theoretic value is that results conditional on RH or Dirichlet GRH could drop the corresponding premise. Their other assumptions would still need to be checked individually. Beyond that, a readable account could show whether the iterated theta lifts, divisibility and error control yield a reusable argument, and identify which steps need the technical machinery of the release. Alpöge expects a substantially shorter proof. That expectation awaits further work. Bombieri’s official problem description · Alpöge’s note

The public record gives AI researchers a testable question. Can people and a separate set of tools reuse, revise or reject the steps proposed by the model? Progress in one closely supervised collaboration with formal verification leaves the answer for other problems and fields open. Alpöge’s note · Sutherland’s verification log

In this series

Stay updated

Get weekly research updates, market signals, and listing intelligence — follow along on Telegram or X.