Daily · Article 01 · 3 min read
- Anthropic
MIT researcher rebuilds Riemann proof files and tests their scope
November 22 replication report and November 23 JST status update
Summary
MIT mathematician Andrew Sutherland reproduced the same proof file in a separate setup and reported that two independent checkers accepted it. Here is what they checked and what Clay lists.
Key facts
MIT mathematician Andrew Sutherland independently checked the proof files that Levent Alpöge said he found with Claude. The hypothesis is a mathematical question tied to how prime numbers are distributed. Sutherland's report, updated at 01:20 JST on November 23, says a rebuild in an isolated environment produced a verification file identical byte for byte to the earlier export. Two independently written checkers also accepted the full exported proof, according to his report. Sutherland's verification records
The November 22 article covered a University of Waterloo researcher's checks. The new report adds evidence about reproducing the file from the same source and about agreement between checkers. The Clay Mathematics Institute's official page still displayed “Unsolved” at the time of review. Official problem page

1. Rebuilding the files and comparing the results
Lean is a language and a tool for writing down and checking mathematical definitions and proofs. The published source holds the proof's ingredients. An export bundles a record that other checkers can read. Sutherland rebuilt the files on a separate machine, with networking disabled and the source mounted read-only. The two exports were identical in content. Each was 61,570,595,788 bytes, about 61.57 GB in decimal units.
The separately stored compiled files matched for 175,186 of 175,187 modules. In the remaining file, parallel processing changed an internal variable name, and that module passed the proof checks. The export and the individual compiled files have their own comparison results.
%% Rebuilding from the published source · editorial diagram
%%{init: {"htmlLabels": false}}%%
flowchart TD
A["Start · published proof source"] --> B["Rebuild · separate machine"]
B --> C["Compare export · earlier file"]The cat asks Q&A
- The researcher
If both files are the same size, can we assume they have the same contents?
Pocket NotebookCompare the bytes themselves. The three-character files ABC and ABD are the same length and differ at the last character. The public logs give the file fingerprint, called a hash, and the result of the byte-for-byte comparison.
2. What the checkers examined
A proof checker checks whether the steps follow from the definitions and assumptions it was given. Imagine checking a spreadsheet claim that the total for all orders is correct. The addition can be sound, while a sheet that excludes cancelled orders changes the scope of the claim. A separate comparison looks at which statements a mathematical proof actually establishes.
| What was checked | Sutherland's reported result | What this check establishes |
|---|---|---|
| Full exported proof | Accepted by nanoda and con-leche | Two independent implementations accepted the logical steps in the export |
| Matching the statements | con-leche and its Rust port con-ron matched five main theorems | The supplied proof supports the statements set out in the challenge |
| Compiled modules | lean4lean accepted all 175,187 | Five of these passed after the check was rerun with a higher step limit |
- What was checked
Full exported proof
- Sutherland's reported result
- Accepted by nanoda and con-leche
- What this check establishes
- Two independent implementations accepted the logical steps in the export
- What was checked
Matching the statements
- Sutherland's reported result
- con-leche and its Rust port con-ron matched five main theorems
- What this check establishes
- The supplied proof supports the statements set out in the challenge
- What was checked
Compiled modules
- Sutherland's reported result
- lean4lean accepted all 175,187
- What this check establishes
- Five of these passed after the check was rerun with a higher step limit
The cat asks Q&A
- The researcher
Did giving five checks more steps add extra mathematical assumptions?
Pocket NotebookThe report says it raised the maximum number of checking steps. The checker had hit its default limit and read the same modules again with more steps available. The separate assumption audit found the five main theorems depended on three standard Lean axioms.
3. Reading “complete” next to the name of the check
The latest MIT summary lists completed checks alongside work in progress. Replaying the original compiled modules with Lean and replaying the full exported record with the Lean checker are separate jobs. The 01:20 update says Lean had accepted every compiled module and describes the pending stage as a repeat using the export. To read a result, start with the file and checker named in its row.
| Work | Confirmed status | Evidence to read next |
|---|---|---|
| Separate rebuild and two independent checker runs | Reported complete | File comparison and checker logs |
Lean replay of the export and lake comparator | In progress | MIT verification updates |
| Clay Mathematics Institute problem status | Unsolved | Official problem page |
- Work
Separate rebuild and two independent checker runs
- Confirmed status
- Reported complete
- Evidence to read next
- File comparison and checker logs
- Work
Lean replay of the export and
lake comparator- Confirmed status
- In progress
- Evidence to read next
- MIT verification updates
- Work
Clay Mathematics Institute problem status
- Confirmed status
- Unsolved
- Evidence to read next
- Official problem page
The cat asks Q&A
- The researcher
If one more check passes, can I read that as “the Riemann hypothesis is solved”?
Pocket NotebookRead which check finished. A machine check tests whether a given statement follows from the stated assumptions. The mathematical community's review can build on that result. Millennium Prize rules
To close
What makes a verification report reusable
A report that names the file, checker and checked statement gives another team a way to repeat the work. Sutherland says he used Claude Code for the checks. The figures and outcomes here come from his public logs.
Sources
- MIT verification updatesmath.mit.edu
- Official problem pageclaymath.org
- Millennium Prize rulesclaymath.org