AI agents attack open problems in mathematics. A result counts only when it is a Lean 4 proof the computer has checked, so it can be trusted without trusting the AI or the search code behind it.
| Date | Problem | Solution | Proof | Paper |
|---|---|---|---|---|
| 2026-11-05 – 2026-11-07 | Signed difference sets | 16,101 open entries of Gordon's table settled: 50 sets found, 16,051 proved impossible by about twenty Lean-checked rules | solutions.md | P2 (draft) |
| Problem | The question | Status |
|---|---|---|
| Signed difference sets | Which ±1 patterns with perfectly flat autocorrelation exist? Settling open entries of Gordon's table. | 16,101 open entries settled: 50 sets found, 16,051 proved impossible; 91% of Gordon's table settled, counting results from the literature |
| Catch-Up | Is the number game Catch-Up on 1..N always a draw when the total is even? | open; partial results |
| R(K₃, K₄−e, K₄−e) | Is this three-colour Ramsey number 21 or 22? | open; search engine being built |
All candidate problems, rated: problems/problems.md.
- Lean 4 checks every proof: no
sorry, nonative_decide, and#print axiomsshows onlypropext,Classical.choice,Quot.sound. - A separate agent tries to break each result before it counts.
- Work by others is credited; their results are not listed as ours.
Check it yourself (Lean 4 and Mathlib v4.35.0), from the repo root:
lake exe cache get # prebuilt Mathlib, once
lake build --wfail SignedDifferenceSets # one problem's library
A fresh build takes hours and much RAM; each problem's README says what it needs.
A problem is chosen from the rated list and stated in Lean. An attack loop then searches, conjectures and proves; agents bring fresh ideas or try to refute results. Details: CLAUDE.md, principles.md.
In Claude Code, from the repo folder:
/choose: pick a new problem and set it up./solve <problem>: work on a problem (its folder name), until you stop it (Esc)./maintain: work on the repo itself (workflow, skills, templates); never inside/solve.
- Code and Lean proofs: Apache-2.0.
- Data and text (problem tables, objects, documentation, papers): CC BY 4.0.
- Imported files keep their own license and notice, next to them (e.g.
problems/catch-up/lean/CatchUp/BoCowgill/); data from others is credited in the problem'sNOTICE.md(e.g. Gordon's SDS data).