What has changed since the announcement
Mathematician Levent Alpöge announced on November 20, 2026, that he had developed a proof of the Riemann hypothesis with help from Claude. He published a short note and a large formal archive written in Lean. As of November 22, the key development is that independent formal checking has moved well beyond the initial compilation tests: Andrew Sutherland, a mathematician at MIT, reports that several checks of the files and main theorem statements have been completed.
The conclusion needs a precise distinction. The checks support that the theorems encoded in Lean follow from the standard axioms used by the system, and that the main statements match targets written independently for review. That is not the same as a readable mathematical explanation, nor does it determine how much Claude contributed. It also does not, by itself, make the work a result already understood and accepted by the entire mathematical community.
The date matters because the status changed quickly. On November 20 and 21, some end-to-end checks were still in progress; Sutherland’s report, updated November 22 at 03:10 UTC, records later results: every module compiled, Lean’s kernel replayed the modules, and several independent checkers accepted the exported file. In that update, the report still lists two Comparator processes—the full run and Lean-kernel replay of the exported proof—as ongoing.
What the proof claims and what it covers
The Riemann hypothesis concerns the nontrivial zeros of the Riemann zeta function. Each zero corresponds to a complex number; the conjecture says that the real part of every one is exactly one half. The importance of that location is tied to how far the count of prime numbers can deviate from its expected approximation. Numerical checks can examine many zeros, but the conjecture applies to all of them, at every height.
Sutherland’s report identifies five central theorems and compares their statements with an independent specification. The first uses the definition of the Riemann hypothesis already contained in Mathlib, Lean’s mathematical library. Another states the same condition directly for the zeta function. The remaining three concern zeros of Dirichlet L-functions, which help mathematicians study patterns in primes across residue classes.
The label “generalized Riemann hypothesis” has a specific, limited scope here. The checked theorems cover Dirichlet L-functions for the moduli and characters described in the report. They do not automatically cover other families sometimes included under broader uses of “generalized hypothesis,” such as Dedekind zeta functions or Hecke and automorphic L-functions. The scope must be read from the formal statements, not inferred from the broad name of the problem.
Scope of the five formal statements
| Statement | What the report establishes | Scope limit |
|---|---|---|
| Classical Riemann hypothesis | The nontrivial zeros of the zeta function have real part equal to one half, using Mathlib’s definition. | Concerns the Riemann zeta function. |
| Explicit version for the zeta function | States the same condition in another formal statement, which Lean recognizes as definitionally equivalent. | Not an independent theorem about a different function. |
| Dirichlet L-functions | Rules out zeros to the right of the critical line under the stated conditions and places zeros in the critical strip on that line. | Covers the specified types of moduli and characters; the pole case is handled by an exception. |
| Primitive-character version | States the condition for primitive characters, the usual form of the generalized hypothesis for this family. | Does not extend the claim to every family of L-functions. |
What independent verifiers checked
The original archive contains 175,187 Lean source modules. The MIT report says that all compiled without errors and that Lean’s kernel replayed all 175,187 modules. The axiom audit for the five main theorems found only the three standard axioms listed in the report: `propext`, `Classical.choice`, and `Quot.sound`. The report also says that no axioms were declared outside Lean’s core.
The checks do not rely on a single program. The exported proof file was accepted by `nanoda`, an independent kernel implemented in Rust, and by `con-leche`, another independent implementation. `con-leche` and its Rust port, `con-ron`, also compared the solution’s five theorems against challenge statements written separately and confirmed that all five were justified. In addition, `lean4lean` accepted all 175,187 compiled modules; five had to be rerun with a higher step limit after reaching the default limit, and then passed.
A separate team rebuilt the project in an isolated environment, with no network access and the source code set to read-only. That rebuild produced an exported file identical, byte for byte, to the published one. The report also describes a reproducibility exception at the compiled-file level: one module produced two possible outputs because of nondeterministic compilation in the version used. The differences affected the name of a bound variable, not the kernel’s result; a later Lean version fixed the behavior described.
These checks address concrete risks: that the project might fail to compile, that the theorem might rely on undeclared extra axioms, or that the exported file might not correspond to the announced mathematical targets. Lean checks a formal chain of deductions against the definitions and axioms it is given. Interpreting those definitions, understanding the mathematical strategy, and explaining why it works remain tasks for mathematical analysis.
What formal verification means
- 01The proof system receives a formal mathematical statement and its definitions.
- 02The verifier checks that each step follows from earlier steps and the permitted axioms.
- 03Independent verifiers repeat or cross-check the process using different implementations.
- 04Mathematicians interpret the statement and study the strategy to determine what the result means and how it relates to the original problem.
What we know—and do not know—about Claude’s role
Alpöge credits Claude with a decisive contribution in the final stage of the research and describes a collaboration in which he directed the investigation. His note refers to weighted sums related to primes and a chain of constructions called theta lifts. That account is the researcher’s description of how the work developed, not an independent measurement of each contribution.
The primary material described in the verification report confirms which files were published and which statements were formalized. It does not identify a specific Claude version or break down which mathematical steps were proposed by the model, which were directed by Alpöge, or how intellectual authorship was divided. The success of the formal checks therefore does not quantify the AI’s contribution or reconstruct the collaboration process.
Formalization also does not replace a comprehensible mathematical exposition. A formal file can be very large and still allow a verifier to check its deductions mechanically. An explanatory paper would serve a different purpose: present the ideas, justify the strategy, and help specialists assess its relationship to earlier results. Mechanical validation and understanding a proof are complementary, not interchangeable.
Why the result matters—and what limits remain
A correct proof of the classical hypothesis would turn into theorems mathematical results that currently depend on it as an assumption. The conjecture gives especially precise control over fluctuations in prime counting, but that theoretical consequence should not automatically be translated into an immediate practical advance in cryptography or a guarantee about specific applications. The consequences depend on each result and the conditions under which it is used.
The verification report adds an indirect arithmetic check: when the proof is combined with an independent formalization of Robin’s and Lagarias’s criteria, it yields inequalities about the sum of divisors. Verifiers also checked the resulting statements and their axioms. This shows that the result connects to mathematical consequences formulated without directly mentioning the zeta function in those two theorems; it does not replace scrutiny of the hypothesis’s formalization or turn every conditional application into a technological benefit.
The formal scope is clearly stated: the checked declarations include the classical hypothesis as defined in Mathlib and a generalized version for Dirichlet L-functions. Mathematical acceptance is a separate matter. The proof has passed substantial formal checks, but explaining its ideas and evaluating it across the mathematical community are different from merely running verifiers. **Formal checking supports the stated theorems; it does not by itself settle questions of authorship, understanding, or academic recognition.**
How to describe the claim’s status
As of November 22, 2026, it is no longer accurate to say that the proof has no independent checking: MIT’s updated report documents full compilation, Lean-kernel replay of the modules, and acceptance of the exported file by several verifiers. It would also be inaccurate to present verification as a complete understanding of the argument: the technical report certifies specific properties of the formal artifact and lists some Comparator processes that were still running in the stated update.
The most precise description is this: the submission’s five main theorems have passed several independent formal checks, including the classical hypothesis and a generalized version restricted to Dirichlet L-functions. The attribution of the work to Claude comes from Alpöge; the verification materials do not identify the specific model or quantify its contribution. Assessing the ideas and their mathematical reception is not reducible to the verifiers’ output.
When evaluating future claims about formal proofs, it helps to distinguish three levels. First, compilation and mechanical checking of the steps. Second, whether the formal statement matches the problem being claimed. Third, mathematical scrutiny that makes the strategy understandable, evaluates its relation to prior knowledge, and identifies its limits. In this case, MIT’s records document firm progress on the first two checks for the published statements; the third requires mathematical reading and analysis, not computation alone.
Three levels that should not be confused
| Level | Question | What it can establish |
|---|---|---|
| Mechanical checking | Does the project compile, and do verifiers accept the formal deduction? | That the encoded steps pass the stated checks under the specified axioms and tools. |
| Statement correspondence | Does the formal theorem match the declared mathematical problem? | That the compared targets correspond to the published statements, within the stated scope. |
| Mathematical understanding | Is it clear why the strategy works and how it fits into the theory? | An explanatory and critical assessment of the ideas; this does not follow automatically from compilation. |
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction