A mathematical claim becomes much easier to assess when the public record separates the theorem, the proof, and the software that checked it.
The Lean and Mathlib communities issued a joint statement on November 21, 2026, certifying Levent Alpöge’s formal proof of the Riemann Hypothesis. Their statement, endorsed by the Lean FRO and Mathlib Initiative, says the proof establishes the intended mathematical statement.
This advances the story beyond the initial announcement. Coverage now needs to explain what was verified and identify the exact scope. It must also distinguish completed checks from those still running.
MustHave.ai has reviewed the public statements and verification record. We have not rebuilt the proof, independently audited its mathematics, or run the checkers.
What changed after the first announcement
The community statement describes a library of about 17 million lines released on November 20. It reports successful compilation, agreement with Mathlib’s RH statement, and checks of the exported proof by multiple implementations.
Its strongest contribution is combining those steps. Producing a large repository would establish that code exists. Matching the intended theorem and checking its justification address different questions about what that code establishes.
On November 20, Terence Tao described the AI-assisted effort as developing verification and interpretation. His account also discussed separate fluid-equation results. Those results should not be folded into a single claim that one verification settled every problem mentioned.
Four questions behind a formal-proof claim
The Xena Project’s November 12 explanation sets out four distinct concerns. Does the formal statement express the intended mathematics? Does the code compile? Does it prove that statement without changing the assumptions? Does the proof survive checking beyond a potentially flawed implementation?
These questions explain why a screenshot of successful compilation is incomplete evidence. A computer can accept the syntax of a statement without establishing the statement itself. A proof can also establish a different claim than the one readers think it is discussing.
An everyday analogy is an invoice calculation. A spreadsheet can calculate perfectly while using the wrong quantity. Confirming the arithmetic and confirming the input mean different things. Formal mathematics needs its own precise version of both checks.
The analogy has limits: a proof checker is not an ordinary spreadsheet. Its value here is to explain why checking the target matters alongside checking the derivation. A reader does not need to inspect millions of lines to understand that distinction.
Read the detailed log at the level of individual checks
Andrew Sutherland’s MIT verification record, updated November 22 at 04:35 UTC, reports the following results. This is a dated snapshot of his work, not a new verification by MustHave.ai.
| Question | Reported result |
|---|---|
| Can the source build? | All 175,187 modules compile without errors. |
| Can another build reproduce the export? | A separate sandboxed build produces an identical export. |
| Do independent implementations accept it? | Nanoda and con-leche accept the complete exported proof. |
| Do separate comparison tools match the theorems? | Con-leche and con-ron comparison modes accept all five headline theorems. |
| Has every additional run finished? | Lean export-kernel replays and an end-to-end comparator run remain in progress. |
The log’s GRH scope is Dirichlet L-functions. It does not list a verification covering every Dedekind, Hecke, or automorphic L-function. Expanding “GRH” beyond that stated scope would change the mathematical claim.
A running entry does not cancel a completed independent check. It also should not be rewritten as a completed result. Both can appear in the same accurate account because they describe different runs.
Keep the theorem’s scope attached to every headline
RH concerns the location of the nontrivial zeros of the Riemann zeta function. The familiar assertion places their real part at one half. A claim about some zeros, a proportion of zeros, or a restricted region has a different scope.
That distinction matters for readers encountering several AI mathematics headlines in succession. Our earlier coverage of a 67.2% lower-bound result addressed a different theorem. It should remain identifiable as that earlier result.
Use a short claim record when comparing announcements. Write down the object being studied, the range of cases, and any assumptions. Then compare the record with the formal target identified by the people doing the verification.
For example, “a result about a proportion of zeros” and “the full RH statement” cannot share a checkmark merely because both mention zeta. The label on a verification table needs to name the claim it belongs to.
Formal acceptance and explanation serve different readers
The community statement says further work will help explain the proof’s ideas. That is a useful direction for the next stage of coverage. Verification supplies evidence that a formal derivation works; exposition helps people understand and reuse the argument.
A working mathematician may want to know which intermediate idea changes the subject. A developer may want a reproducible artifact and a clear trust boundary. A general reader may want an explanation of why several checkers increase confidence.
Those questions deserve different evidence. A prose summary can explain an argument while omitting details needed to rebuild it. A successful rebuild can establish reproducibility while offering little intuition about why the theorem is true.
When a readable account appears, compare its theorem and assumptions with the verified formal target. Do not assume that a later document automatically inherits the status of an earlier artifact. The correspondence is another claim to examine.
A practical way to read the next verification update
Keep a small record with an artifact identifier, a theorem description, a checker version, and a result. Add the time and a link to the actual output. This proposed record prevents a summary from silently mixing different releases.
Separate three possible changes. A new checker may finish on the same export. The author may replace the proof archive. An explanatory document may change without changing the formal artifact. Each change calls for a different story update.
If the export changes, keep earlier acceptance attached to the earlier export until the new one is checked. If only another run finishes, preserve the prior successes and add its result. If a prose explanation changes, identify which argument or interpretation changed.
This approach also makes corrections manageable. Readers can follow one specific claim back to one specific record. The size of a repository or the reputation of an AI model cannot replace that chain.
What readers can conclude from the current record?
The community has publicly certified the RH statement and formal proof. The detailed log documents independent successes and names additional work still running. The published mathematical scope should stay attached to any account of those results.
The next useful reporting can examine explanations of the proof’s ideas and material changes to the verification record. A prize decision, a broader theorem, or a review of a separate manuscript would need its own evidence.
Read the community statement alongside the detailed MIT log, then use the Xena explanation to understand why the checks are separate. Our Fermat formalization coverage provides another example of asking what a proof package establishes.
Which part needs a clearer explanation: the formal statement, the independent checkers, or the step from a checked derivation to mathematical understanding?



