AI-Assisted Lean Proof Of Riemann Hypothesis Clears Multiple Independent Kernel Checks
MIT mathematician Andrew Sutherland announced Friday in Cambridge that Levent Alpöge's computer-formalized proof of the Riemann hypothesis has cleared a battery of independent checks inside and outside the Lean proof assistant, sharply boosting confidence that the AI-assisted argument is logically sound.[1]
In a verification report posted Friday, Sutherland said all 175,187 Lean modules in Alpöge's project compiled without errors under Lean v4.33.1 with the matching Mathlib library, and kernel replays using Lean's own checker, a Rust proof kernel called nanoda, and the third-party con-leche tool each accepted the exported proof file corresponding to the main Riemann hypothesis theorems.[1] He also documented that the key theorems rely only on Lean's standard small set of foundational axioms, with no hidden assumptions added by the project.[1]
On Friday Sutherland reported that a full leanchecker kernel replay of every module in the project finished without a single rejection at about 12:17 pm Central, confirming that Lean's own trusted core can reconstruct all in-IDE proofs from scratch.[1] The exported proof file "solution.ndjson.zst" from the project's Main module was hashed in both compressed and uncompressed form and shown to match the challenge export published earlier by the Deicyde/riemann-hypothesis continuous integration system, which is meant to prevent any tampering or version mismatch between what was proved and what is being verified.[1] Later that afternoon, the nanoda Rust kernel accepted the full exported proof, exiting with status 0 at 1:05 pm Central after running for just under 28 minutes on an 80-thread machine, while con-leche, in its most stringent "--verified" mode, validated more than 1.46 million declarations in a little over three minutes on 64 threads.[1]
Additional cross-checks focused on making sure the external verifiers were confirming the right theorems and not relying on any undocumented logical power. In a "--compare" mode run, a special branch of con-leche checked that all five Riemann-hypothesis-related "challenge theorems" posed by the Deicyde competition were indeed justified by Alpöge's submitted solution, accepting more than 51,000 challenge declarations before exiting successfully.[1] Sutherland used Lean's "#print axioms" command and an audit log to show that these headline theorems depend only on three standard axioms built into Lean's classical mathematics mode, with a total of 15 axioms in the environment and no extra assumptions declared in the project itself, addressing a central concern that a supposed proof might secretly assume the conclusion it is trying to show.[1]
A separate tool, lean4lean, rechecked all 175,187 compiled modules as a further guard against subtle bugs in the main Lean binary, succeeding on every file; five modules that initially hit deterministic timeouts on intensive decision or kernel computations passed cleanly once their step limits were raised, suggesting performance rather than correctness issues.[1] Some older bundled versions of the con-leche and con-ron checkers that ship with Lean v4.36.0 refused to process the export because they lack support for a reflexive inductive type used in Alpöge's construction, and the lean4lean bundled with that newer Lean release could not read the older v4.33 core library, but an updated con-leche master branch handled the proof without trouble.[1] As of 4:05 pm Central Friday, Sutherland noted that additional end-to-end runs by the lake comparator and Lean's "default" and "paranoid" kernels were still underway, although an earlier comparator run with Lean v4.36.0 had already passed both statement comparison and the axiom check once.[1]
The verification wave follows twin Scientific American reports Friday morning that described Alpöge, working with Anthropic's AI tools and the Lean theorem prover, as claiming a proof of the Riemann hypothesis and quoted independent mathematicians who, after reading human-level outlines and machine-checked fragments, judged the argument almost certainly correct in broad structure.[2][3] Those articles situated the work within a new model of collaboration in which human mathematicians guide automated reasoning systems and then rely on independent proof kernels to rule out both human slips and software bugs, a model Sutherland's multi-kernel audit was designed to stress-test on what has long been regarded as the most famous unsolved problem in mathematics.[3]
Show source details & analysis (3 sources)
📌 Key Facts
- On November 20, 2026, MIT mathematician Andrew Sutherland reported that all 175,187 Lean modules in Levent Alpöge's Riemann hypothesis formalization compile without errors under Lean v4.33.1 with Mathlib v4.33.1, according to the posted RH Lean proof: verification files.
- Sutherland confirmed using `#print axioms` and an audit log that the five headline theorems in Main.lean depend only on Lean's three standard axioms (propext, Classical.choice, Quot.sound), with no extra axioms beyond Lean's core and a total of 15 axioms in the environment documented in the RH Lean proof: verification files.
- A leanchecker kernel replay of every module succeeded for all 175,187 modules with zero rejections, completing around 18:17 UTC on November 20.
- The exported proof file `solution.ndjson.zst` from module Main has specified compressed and uncompressed SHA-256 hashes and matches the challenge export published by Deicyde/riemann-hypothesis CI, as recorded in the RH Lean proof: verification files.
- The independent Rust proof kernel nanoda accepted the full exported proof on November 20, exiting with status 0 at 19:05:45 UTC after a 27 minute 42 second run on 80 threads.
- The con-leche proof checker in "--verified" mode (master branch cd2eceab) validated 1,468,418 declarations on the full export and exited successfully at 21:19:04 UTC after 3 minutes 15 seconds on 64 threads, according to the RH Lean proof: verification files.
- In `--compare` mode (branch compare, commit 5dc36dc1), con-leche verified that all five challenge theorems are justified by the submitted solution, accepting 51,692 challenge declarations and exiting with status 0 at 21:19:12 UTC.
- The lean4lean tool (commit 67f63b0c, built with Lean v4.33.1) successfully rechecked all 175,187 compiled modules, with five modules that initially hit deterministic timeouts on decide/kernel computations passing when rerun with higher step limits.
- Older bundled versions of con-leche and con-ron shipped with Lean v4.36.0 declined to process the export (exit code 2) because they do not support a reflexive inductive type used in the proof, while the updated con-leche master could; the lean4lean bundled with Lean v4.36.0 could not read the v4.33 core library.
- As of 22:05 UTC on November 20, 2026, lake comparator and the Lean default and paranoid kernels were still running end-to-end checks on the export, although an earlier comparator run with Lean v4.36.0 had already passed statement comparison and the axiom check once.
📰 Source Timeline (3)
Follow how coverage of this story developed over time
- On November 20, 2026, MIT mathematician Andrew Sutherland reported that all 175,187 Lean modules in Levent Alpöge's Riemann hypothesis formalization compile without errors under Lean v4.33.1 with Mathlib v4.33.1.
- Sutherland confirmed via `#print axioms` and an audit log that the five headline theorems in Main.lean depend only on Lean's three standard axioms (propext, Classical.choice, Quot.sound) and that no additional axioms are declared outside Lean's core, with 15 axioms total in the environment.
- A leanchecker kernel replay of every module succeeded for all 175,187 modules with zero rejections, finishing at about 18:17 UTC on November 20.
- The exported proof file solution.ndjson.zst from module Main has compressed SHA-256 hash 61a8938ad5f5e57397c1c0e3b954916af546d9004ade01c3555744a681b05381 and uncompressed SHA-256 hash 5ea74b912147c5c93cc2b6d70cc727cf908df2f62ac6a20480a3fa02f6be244e, and matches the challenge export published by Deicyde/riemann-hypothesis CI (SHA-256 bd484ac9886a96f26b1d9a7a59ebd45b7a70d102c015d4f31f9302bbdd6ab2b9).
- The independent Rust proof kernel nanoda accepted the full exported proof on November 20, exiting with status 0 at 19:05:45 UTC after a 27 minute 42 second run on 80 threads.
- The con-leche proof checker (master branch cd2eceab) validated 1,468,418 declarations in "--verified" mode on the full export, exiting successfully at 21:19:04 UTC after 3 minutes 15 seconds on 64 threads.
- In --compare mode (branch compare, commit 5dc36dc1), con-leche verified that all five challenge theorems are justified by the submitted solution, accepting 51,692 challenge declarations and exiting with status 0 at 21:19:12 UTC.
- The lean4lean tool (commit 67f63b0c, built with Lean v4.33.1) successfully rechecked all 175,187 compiled modules; five modules initially hit deterministic timeouts on decide/kernel computations but passed when rerun with higher step limits.
- Older bundled versions of con-leche and con-ron shipped with Lean v4.36.0 declined to process the export (exit code 2) because they do not support a reflexive inductive type used in the proof, while the updated con-leche master could; the lean4lean bundled with Lean v4.36.0 could not read the v4.33 core library.
- As of 22:05 UTC on November 20, 2026, lake comparator and the Lean default and paranoid kernels are still running end-to-end checks on the export, although an earlier comparator run with Lean v4.36.0 has already passed statement comparison and the axiom check once.