The defensible answer is narrower than the headline. As of the November 22, 2026 snapshot, two independently written Lean kernels had accepted the full exported proof, and two comparator implementations had matched all five proved statements against the challenge. That is substantial mechanical verification of specific formal statements. It is not independent confirmation that AI authored the proof, and two checks, including replay through Lean’s own kernel, were still in progress. The practical consequence is straightforward: cite the checker results, scope and pending work, rather than repeating “AI proved the Riemann Hypothesis” as a settled fact.
The available evidence also has sharp boundaries. It includes Andrew Sutherland’s verification page and several community mirrors and verification harnesses, but not a direct examination of the announcement PDF, expert mathematical commentary on the proof’s significance, or outcomes for the two unfinished checks. The AI-authorship claim comes from the announcement as described by an unmodified mirror, whose own title says “Not independently verified.”
What was announced, and what is actually attributed to AI
The verification page records that the Lean proof was announced at alpo.ge/remark.pdf on November 20, 2026. It credits Andrew Sutherland of MIT and says the checks were run with Claude Code. The mirror describes the accompanying note as claiming “a proof of the Riemann Hypothesis, and of GRH for Dirichlet L-functions, found working with Claude,” presented as Lean code.
Those are three different propositions:
- A Lean development containing the stated proofs exists.
- Independent software accepted aspects of that development.
- The proof was found while working with Claude.
The published checks address the first two. They do not independently establish the third. Nothing in the verification results can reconstruct the interaction, determine how much of the mathematics originated with the model, or corroborate the announcement’s account of its discovery.
That distinction matters because mechanical verification does not depend on resolving the authorship dispute. A person, an AI system or an unknown mixture of both could produce the artifact; the checkers still evaluate the same exported proof. But a report about the artifact should not quietly turn the announcement’s attribution into a verified biographical fact.
The verification stack, layer by layer
The published results form a chain. A rebuild ties the checked output to a particular source archive. Compilation establishes that the source builds. Kernels check the exported proof. Comparators check which statements were proved. An axioms audit inspects the assumptions on which those theorems depend.
The following table separates the reported results from their limits. The failure-mode explanations are an interpretation of what each layer can establish, not additional claims by the verifiers.
| Layer | Reported result | What it establishes | What it does not settle |
|---|---|---|---|
| Artifact identity | A 14-part archive was downloaded on November 20, joined and extracted unmodified; the joined archive’s SHA-256 is published by the reproduction mirror. | Others can identify and compare the same archive. | A hash establishes artifact identity, not mathematical correctness. |
| Source-tree audit | The automated verifier audits 175,187 .lean files, roughly 1.4 GiB, for prebuilt .olean or binary artifacts and scans the entry files for tokens including sorry, admit, axiom, opaque, implemented_by, @[extern] and native_decide. | The listed placeholders, extensions and smuggled binary artifacts are covered by explicit checks. | A token scan is not a general proof that every conceivable supply-chain attack is impossible. |
| Compilation | All 175,187 modules compiled with zero errors under Lean v4.33.1 and Mathlib v4.33.1, according to Sutherland’s page. | The development builds under the pinned toolchain. | Successful compilation alone is not acceptance of the exported proof by a kernel. |
| Independent kernel acceptance | The independently written kernels nanoda and con-leche accepted the full exported proof. | Acceptance does not rest on one checker implementation alone. | The result still needs statement matching to establish what was proved, and the first-party Lean-kernel replay remained pending. |
| Statement matching | The --compare modes of con-leche and its Rust port con-ron, independent implementations of Comparator’s statement matching, each confirmed all five theorems against the challenge. | The checked statements were matched to the challenge by two implementations. | Matching does not establish mathematical importance or extend the result to statements outside the challenge. |
| Axioms audit | The independent verification repository reports only propext, Classical.choice and Quot.sound. | The checked theorems do not depend on additional reported axioms beyond Lean’s three standard classical axioms. | A clean audit does not show that a theorem is strong, surprising or mathematically significant. |
| Sandboxed rebuild | An independent sandboxed lake build reproduced the export byte for byte. | The exported artifact was reproducible from the archive in that environment. | Byte identity does not replace proof checking or establish the announcement’s authorship account. |
| Remaining runs | The lake comparator run and Lean-kernel replay of the export were still in progress in the November 22 snapshot. | The page explicitly identified unfinished work rather than presenting the stack as closed. | Their eventual success or failure cannot be inferred from the other completed checks. |
Compilation is not the load-bearing result
It would be easy to summarize the development as “175,187 files compile,” but that is the weakest interpretation of the announcement. A large source tree can build successfully without telling a reader whether the intended theorem was proved, whether the proof export survives independent checking, or whether an additional axiom slipped into the dependency chain.
The stronger results are kernel acceptance and statement matching. The two independently written kernels accepted the full exported proof, while the two comparator implementations confirmed all five statements against the challenge. The mirror provides useful context for that target: src/Statements.lean imports only Mathlib and states GRH for Dirichlet L-functions, while src/Main.lean proves riemannHypothesis : RiemannHypothesis, using Mathlib’s own definition, plus four statements about Dirichlet L-functions.
This is also where the word “independent” needs discipline. The evidence supports independence between software implementations: two kernels and two comparator implementations. It is not equivalent to broad independent scrutiny by mathematicians, and the supplied material contains no expert assessment of the proof’s novelty, significance or relationship to existing mathematics.
The axioms audit answers a different question again. The verification harness runs #print axioms on all five theorems and fails if anything outside propext, Classical.choice and Quot.sound appears. It explicitly rejects sorryAx, Lean.ofReduceBool, Lean.trustCompiler and custom axioms. This guards against accepting a theorem whose apparent proof depends on an undeclared shortcut or extra assumption. It does not judge whether the formal statement captures everything an informal headline might suggest.
Claude Code ran checks; deterministic kernels did the accepting
Sutherland’s page says the checks were run with Claude Code. The kernels and comparators supplied the reported acceptance results. Those roles should not be collapsed into “Claude verified the proof.”
An orchestration agent can coordinate builds, invoke tools and collect results, but the evidential weight here comes from deterministic checkers accepting an artifact. If an agent merely reported that a proof looked correct, there would be no comparable acceptance layer. Conversely, agent orchestration does not invalidate a deterministic result; it makes the underlying checker outputs and reproducible artifacts the things to inspect.
For engineering teams using agents in verification work, this is the reusable distinction. Preserve the exact artifact, the tool versions and the checker outputs. Treat agent narration as a report about the process, not as the proof certificate. The same principle applies to the authorship claim: “found working with Claude” describes the announced discovery process, while kernel acceptance describes a mechanical property of the resulting Lean export.
“GRH” has a narrower scope than the headline suggests
The verification page is explicit: “GRH” here means Dirichlet L-functions, covering all moduli and all characters. Dedekind zeta functions, Hecke L-functions and automorphic L-functions in general were not among the checked statements.
This is not a cosmetic footnote. A statement such as “AI proved RH and GRH” invites a broader reading than the checked artifact supports. The accurate scope is the Riemann Hypothesis plus GRH for Dirichlet L-functions. Extending the claim to every family of L-functions would require evidence that this verification stack does not provide.
The two pending checks impose a separate limit. Independent kernels had accepted the export, but replay through Lean’s own kernel had not yet been reported complete. The lake comparator run was also unfinished, despite the two completed comparator implementations. These statuses describe the November 22 snapshot only. Anyone citing the work later should recheck the verification page rather than assuming either outcome.
Reproduction helps, but each reproducibility claim has its own job
The mirrors make the artifact unusually concrete. Dimitrios Mitsios’s reproduction instructions identify the archive parts and publish the joined archive’s SHA-256: 9631b4f020199e580ba86178cb56c6e9b09f690c2369a80dc4334aa93f49c7fc. The source mirror identifies Lean v4.33.1 and Mathlib commit 0df444a3, corresponding to Mathlib’s v4.33.1 tag. Sutherland’s page reports the byte-for-byte sandboxed rebuild.
Together, those details reduce ambiguity about what was checked. They also prevent a common reporting error: treating a reproducible artifact as interchangeable with a correct theorem. Reproduction answers “did this input produce these bytes?” Kernel acceptance answers “did this checker accept this proof export?” Statement matching answers “was the proved statement the requested one?” The axioms audit answers “which assumptions does it use?” No single row substitutes for the others.
The acceptance bar for the next AI proof claim
I would apply the same layered bar to any AI-generated proof or formal verification result:
- Bind the claim to an exact artifact. Publish its identity, source and toolchain pins so others are not checking an unnamed or drifting version.
- Separate compilation from proof acceptance. Require a proof kernel to accept the exported artifact, with independently written checkers where available.
- Match the formal statement to the challenge. A valid proof of a different or narrower statement does not establish the advertised claim.
- Audit axioms and prohibited shortcuts. Report the accepted axiom set and scan for relevant placeholders, prebuilt artifacts and unsafe entry points.
- Rebuild in a sandbox and compare outputs. Use reproducibility to support artifact integrity without presenting it as mathematical validation.
- Publish scope and unfinished checks together. Pending jobs, excluded statement families and absent expert review belong beside the result, not in a later correction.
For this announcement, that bar supports a careful statement: a Lean proof announced as developed with Claude passed several independent mechanical checks for the Riemann Hypothesis and Dirichlet GRH, with two verification runs still pending in the November 22 snapshot. It does not support independent confirmation of AI authorship, a broader GRH claim, or a judgment about the proof’s mathematical significance. Mechanical verification shifts trust away from the producer’s identity and toward inspectable checkers, but only when the statement, assumptions, artifact and checker status are reported with it.
Frequently Asked Questions
What is the specific scope of the GRH claim in the verified Lean proof?
The verification page is explicit: “GRH” here means Dirichlet L-functions, covering all moduli and all characters. Dedekind zeta functions, Hecke L-functions and automorphic L-functions in general were not among the checked statements.
Which axioms are permitted in the verified theorems?
The independent verification repository reports only propext, Classical.choice and Quot.sound.
Join the discussion
Share a useful perspective or ask a question about this article.