Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

88 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Solving Open Math Problems

A research system in which many parallel Devin sessions hunt for and settle obscure but genuinely open problems in graph theory, spectral graph theory, combinatorial designs, and number theory. Candidate problems are curated from the literature, attacked by independent solve sessions with different strategies, adversarially reviewed, and — where possible — verified down to machine-checkable certificates (Lean 4 kernel proofs and/or DRAT-checked SAT proofs). Settled results are packaged into self-contained handoff folders for human researchers.

Results so far

ID Problem (one-liner) Status Verification Links
P02 Brandt/West regular-supergraph statement: every maximal triangle-free graph with δ ≥ n/3 has a regular supergraph via vertex multiplications REFUTED (9-vertex witness + Farkas certificate) Lean 4 (kernel-checked, no sorry) solutions/P02 · handoff/P02 · formalization/P02
P06 Graffiti / WoW 698: L2 norm of negative adjacency eigenvalues ≤ Randić index PROVED TRUE (WoW 129 remains open; exhaustive to n = 11) Lean 4 + adversarial review solutions/P06 · handoff/P06 · formalization/P06
P07 Graffiti 154 (2m·μ² ≤ n³) and 143 (variance of positive eigenvalues ≤ m/μ) REFUTED — but priority belongs to demonstrandum-research (public 2026-06-12, six weeks earlier; see PRIORITY.md). Independent value here: the Lean formalization of the 154 refutation Lean 4 (154, both readings) solutions/P07 · handoff/P07 · formalization/P07
P08 Graffiti 39/40: deviation of the distance matrix ≤ number of positive (39) / negative (40) adjacency eigenvalues PROVED TRUE Lean 4 + adversarial review solutions/P08 · handoff/P08 · formalization/P08
P13 (9,6,1)-perfect Mendelsohn design — smallest open case of the block-size-6 PMD spectrum NONEXISTENT (adversarially confirmed) Lean 4 + DRAT (drat-trim s VERIFIED) solutions/P13 · handoff/P13 · formalization/P13
P14 Three balanced ternary designs left open by the CPro1 campaigns: BTD(14,18;7,1,9;7,4), BTD(12,15;6,2,10;8,6), BTD(12,20;4,3,10;6,4) NONEXISTENT (adversarially confirmed); 4th instance BTD(14,28;8,3,14;7,6) undecided DRAT-certified SAT + independent CP-SAT (machine-verified, not Lean) solutions/P14 · handoff/P14

All other problems (P01–P25, see problems/README.md) are open or in progress; live per-run status is tracked in runs/INDEX.md.

How the pipeline works

  1. Research — survey the literature for obscure open problems with small plausible witnesses (research/), then curate a problem catalog with formal statements and attack plans (problems/, criteria in context/PLAN.md).
  2. Parallel solves — each problem gets up to 5 parallel Devin Ultra sessions, one per prompt variant (direct search, structured construction, SAT/SMT, LP/ILP duality, literature-first reasoning). Protocol: context/METHODOLOGY.md. Logs land in runs/<Pxx>/<variant>/.
  3. Adversarial review — a hostile second session tries to break every claimed result: independent re-verification, statement fidelity against the original source, priority search.
  4. Certificates — results are pushed to machine-checkable form: standalone verify.py scripts, DRAT-checked UNSAT proofs, and Lean 4 kernel formalizations (formalization/).
  5. Handoff — each settled result becomes a self-contained package for a human researcher (handoff/<Pxx>/), including original sources, proofs, certificates, and priority notes.

Directory map

Directory Contents
problems/ Curated problem catalog P01–P25: statements, status, attack plans, prompt variants
runs/ Per-run working logs and code, one folder per problem × variant; status matrix in INDEX.md
solutions/ Headline artifacts per settled problem: proofs, witnesses, certificates, verify.py
handoff/ Self-contained verification packages for human researchers
formalization/ Lean 4 projects machine-checking the settled results
social/ Tweet drafts and result graphics
research/ Literature surveys used to source candidate problems
context/ Project plan, methodology, and initial candidate lists

Verification levels

  • Lean 4 — statement and proof checked by the Lean kernel (no sorry, no extra axioms).
  • DRAT — SAT-solver UNSAT proof independently checked by drat-trim (s VERIFIED).
  • Adversarially reviewed — a separate hostile session independently re-verified the claim, the statement fidelity, and priority.

Accuracy over polish: every claim here links to its certificates, and priority caveats (notably P07) are preserved verbatim in the handoff packages.

About

No description, website, or topics provided.

Resources

Stars

8 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages