The question that was open is the one that closed

Our November 20 coverage separated this claim into two questions. Whether the code compiles was already partly answered that day. Whether the theorem at the top of the code is the Riemann Hypothesis, rather than a weaker or subtly different statement, was not. That second question is the one a formalization can fail silently, and it is the one the certification addresses directly.

The method matters more than the headline. The community added a file that isolates the statement of the Riemann Hypothesis and imports only Mathlib. That isolated statement was then compared against the theorem Alpöge's library actually proves, using two independent implementations of statement matching. Sutherland's ledger records the same result in its own table: con-leche --compare accepted all five theorems as "justified by the solution" at 21:19 UTC, and its Rust port con-ron --compare accepted all five at 23:02 UTC on November 21. The five theorems are the Riemann Hypothesis in Mathlib's definition, the same statement written out in the submission's own words, and three statements covering Dirichlet L-functions.

Readers should keep the scope attached. Sutherland's page is explicit that "GRH" in the checked statements means Dirichlet L-functions, for all moduli and all characters. It does not mean every L-function mathematicians use. A headline that says the generalized hypothesis is settled without that qualifier is saying more than was checked.

What the ledger shows, line by line

Sutherland's page is the most useful document in this story because it does not summarize. It lists each check, its result and its finishing time.

Passed: all 175,187 modules compile with zero errors. An independent rebuild in a sandbox, with no network and read-only source, on a separate machine, finished with exit 0, and the export made inside that sandbox is byte-for-byte identical to the published export, at 61,570,595,788 bytes. The leanchecker replay of every compiled module passed, 175,187 of 175,187, with zero rejections. That module-level replay is a different check from the kernel replay of the full export listed below: one walks the library module by module, the other replays the single exported proof object end to end. The axiom audit for the five headline theorems lists only the three standard axioms. Two independently written kernels, nanoda and con-leche, accept the full exported proof. lean4lean accepts every compiled module, with five modules needing a raised step limit before they passed.

There is also a cross-check that does not depend on Mathlib's definition of the zeta function at all. Combined with an independent formalization of Robin's and Lagarias's criteria, the proof yields two theorems about the divisor sum, stated without any mention of zeta. For that combined export, Sutherland records a completed lake comparator pass at 04:16 UTC on November 22: the Lean paranoid kernel, nanoda and the Lean default kernel all accept, with the verdict "Your solution is okay!"

The two lines that still say "running"

Against that list, the same ledger's summary is plain about what had not finished at its last update: "The lake comparator run and the Lean-kernel replay of the export are still in progress." Both are listed as running, with no verdict yet. The comparator's statement comparison and axiom check have already passed in one run. Its kernel stages, replaying a roughly 62 GB export, are the slow part.

Coverage of the certification has not always kept these line items separate. Scientific American, updated on November 21, reports the certification and writes of a comparator run that Kevin Buzzard had said was still awaited: "That run has now been completed." Several comparator-based checks have indeed completed, including the statement matching and the pass on the smaller arithmetic cross-check. The end-to-end comparator run on the main export is a different line item, and on the ledger's most recent version, timestamped 04:35 UTC on November 22, it is still listed as running. Which run a report means matters, and the ledger is the document that separates them.

None of this reverses the certification. The Lean and Mathlib statement rests on criteria it spells out, and the independent-kernel passes are the strongest part of the case: a proof accepted by checkers written by other people, in other languages, does not rest on any one implementation being right. But "certified by the community that built the tools" and "every named check on the public ledger has finished" are two different claims, and at 04:35 UTC on November 22 only the first one was true.

What has not changed

Two things the certification does not do are worth keeping in view. The Clay Mathematics Institute's problem page still listed the hypothesis as unsolved on November 21, 2026, and Clay recognition is a separate process from a Lean certification, with its own publication and acceptance requirements. And the sources reviewed offer no human-readable exposition of the full proof. The joint statement itself ends by looking forward to "all the activity in the mathematical community that will help explain the new ideas in this proof." A machine-checked proof of 17 million lines can be correct before it is understood, and that situation will outlast the remaining kernel runs, whichever way the headlines compress it.

Sutherland's page is the place to confirm the last step when it lands.

Sources

  1. [1] Lean and Mathlib communities, "Lean verification of the Riemann Hypothesis," joint statement endorsed by the Lean FRO and the Mathlib Initiative, November 21, 2026Read source
  2. [2] Andrew Sutherland, "Verification files for the Lean proof of the Riemann Hypothesis," last updated November 22, 2026, 04:35 UTCRead source
  3. [3] Levent Alpöge, personal note announcing the proof, November 20, 2026Read source
  4. [4] Scientific American, "The AI Riemann hypothesis proof is true, independent mathematicians confirm," published November 20, updated November 21, 2026Read source
  5. [5] ALFW prior coverage, for context: "Levent Alpöge Claims a Riemann Hypothesis Proof With Claude. The Formal Statement Still Needs Review," November 20, 2026Read source