SII MathRSI Team

A self-improving agent harness for mathematical discovery and proof.

MathRSI evolution: base workflow & domain plugins

v0: standalone tools. v1–v5: shared harness releases. v6: hardest-remainder gates, primal–dual coevolution and sparse dual checks across Q₇ E16, C₄-free Q₇ / Q₈ / Q∞ and N(m,r,s).

Select a problem to trace its progress.

Swipe to explore versions →

One block = one result or result family Color identifies the research problem

Columns trace how the shared workflow and domain plugins evolve alongside the mathematical results incorporated at each stage; click any block or callout for its proof, run context and verification status. Erdős 86: Q₇ 304 (construction verified; ≤304 source-reported) · Q₈ 682–683 · Q∞ ≤ 0.5912405943. Q₇: ≥56-edge matchings extend (complete E16). Happy Ending: N(8,6,6) ≤ 61 · ES(9) ≤ 1148 working (m = 7–9). DD D5: +1 / +31 local (+1 / +81 vs dated Comellas table).

Results

Automated mathematical research: stronger bounds, exact answers and reusable proofs.

Illustration · a 2-coloring with no single-color edge

Property BHypergraph two-coloring

Exponentially growing gains for hypergraph coloring

Lower bounds on m(k)m(k)

  1. kk = 535
  2. kk = 667
  3. kk = 7131
  4. kk = 8267
  5. kk = 9543
  6. kk = 101112
  7. kk = 112275
  8. kk = 124664

Uniform theoremAn additive improvement over the stated RS baseline for every k≥12k \ge 12, with a gain that grows exponentially in k.

Why it matters Below each edge threshold, every kk-uniform hypergraph has a coloring with both colors in every edge. The uniform theorem extends stronger guarantees to infinitely many kk.

Method Combine constraints that must hold in the same hypergraph.

Uniform theorem in LeanRead ↗EN
Scope & verification

m(k)m(k) is the fewest edges of a finite simple kk-uniform hypergraph that is not 2-colorable. End-to-end Lean proofs cover m(5) ≥ 33, 34 and the uniform k ≥ 12 theorem; m(5) ≥ 35 has a complete written proof with Lean formalization in progress, and k = 6–12 rest on parameter-specific computer-assisted proofs.

Exact d₂(n,k) values · hatched: still open

Error-correcting codesBinary linear code distances

Closing three code-distance gaps dating to 1985

d2(n,k)d_2(n,k): the largest minimum distance of a binary linear code of length nn and dimension kk

d2(43,10)d_2(43,10)=16

d2(44,10)d_2(44,10)=17

d2(44,11)d_2(44,11)=16

d2(42,16)d_2(42,16)=12†

16≤d2(45,11)≤1716 \le d_2(45,11) \le 17open

† Applies a published 2020 theorem.

Why it matters The main proof closes the first three gaps recorded in 1985. The fourth exact value, d₂(42,16)=12, follows from a published 2020 theorem; d₂(45,11) narrows to 16–17.

Method Reduce every candidate to a complete parent-code domain.

Four exact distancesRead ↗EN
Scope & verification

The first three follow from the original exclusion; 42/16 applies a published 2020 theorem. Lean reductions keep explicit literature, classification and computation premises. d₂(45,11) remains 16–17.

Illustration · Petersen graph, degree 3 and diameter 2

Degree–diameter problemLargest graphs

Our best constructions by setting

Latest project construction in each showcased setting.

D5 follow-up run: v4 + local updates → v5 main maintenance.

Why it matters More vertices under the same degree and diameter limits, with each setting retaining its own best construction.

D5 method Preserve the source for D12; bind the geometric core and check the boundary for D11.

Four showcased settingsResults & methods ↗EN + 中文Historical (10,4) ↗EN
Scope & verification

Both D5 final graphs regenerate deterministically. D12 preserves the 42,402-vertex source and checks all 31 new vertices; D11 takes core–core distances from the H9 theorem under its stated σ/π identification and checks all 161 peripheral and new vertices. End-to-end Lean proofs and independent refereeing for these D5 graphs remain open. On the Comellas table checked 2026-11-16, (10,4) is 2,612 and (16,4) is 16,848, above our earlier constructions of 2,567 and 16,416.

Q₄ with one square (C₄) · density scale

Erdős Problem 86Square-free hypercube subgraphs

Finite and asymptotic bounds for C₄-free cubes

Subgraphs of QnQ_n without C4C_4: per-setting results

  • Q₇ (n = 7)304304-edge construction verified; 304 upper bound as reported in the source package, not re-verified in this pass.
  • Q₈ (n = 8)682–683Four 682-edge constructions + Theorem B (refereeing pending); ≤682 open.
  • Q∞ (density)≤ 0.5912405943T3 proved · T4 ≤ 0.5912405942 · limit 1/2 open.

Why it matters Tighter ceilings in Q₇, Q₈ and arbitrary dimension Qₙ narrow how many cube edges can avoid a four-cycle.

Method Couple local flags, degree potentials and primal–dual coevolution on the same graph.

Computer-assisted bounds · no LeanResults & methods ↗EN + 中文Historical Q₄ ↗中文
Scope & verification

Q₇: two explicit 304-edge C₄-free subgraphs verified in-repo; 304 upper bound as reported in the source package, not re-verified in this pass (21 raw .cnf/.drup traces omitted upstream). Q₈: four 682-edge constructions and C++17 13,150,453-completion replay verified; Theorem B uses 83 active cuts and 25 imported Q₆ DRAT certificates (reviewed in the same model family; independent refereeing pending), and ex(Q₈, C₄) ≤ 682 remains open. Q∞: T3 = 681653995482306581 / 1152921504606846976 ≤ 0.5912405943 and T4 = 9701859764505653085708308080340303 / 16409326186106551516455239680000000 ≤ 0.5912405942 (via a 5-line hypergeometric tower for all n ≥ 1024), both proved and rechecked in 128-bit C++20; exploratory 0.5907 is an approximate fit (residual ≈ 3.9 × 10⁻⁴), and an earlier 0.5899 claim was refuted and withdrawn.

Illustration · one matching (red) inside one Hamilton cycle of Q₆

Ruskey–Savage conjectureHamilton cycles in hypercubes

From Q₆ to dense Q₇ matchings

TheoremIn Q6Q_6, every matching extends to a Hamilton cycle.

CorollaryCombined with published results, for every nn, any matching in QnQ_n that uses at most six directions extends to a Hamilton cycle.

Finite universal certificate · 1024 forests · rank 1024

Why it matters Any set of disjoint edges in the six-dimensional cube can be included in one cycle visiting all 64 vertices. This settles the six-dimensional case; the general conjecture remains open.

Method A finite forest-rank certificate covers the universal matching family.

Computational proof · refereeing pendingRead ↗中文
Scope & verification

The checked replay rebuilds 1,024 forests and all 1,048,576 connectivity bits (rank 1,024), while the larger P6 certificate streams are inventoried rather than replayed here; reviewed in the same model family, with independent refereeing pending and no Lean formalization.

Cup, cap and a convex heptagon

Happy Ending problemErdős–Szekeres & signotope N(m,r,s)

Signotope capacities and planar bounds across m = 7, 8, 9

Ordered rank-3 signotope capacities N(m,r,s)N(m,r,s) & planar ES(m)\mathrm{ES}(m)

  • m = 7 (7-gons)ES(7) ≤ 77Analytic N(7,5,7) ≤ 39, N(7,6,6) ≤ 49 & 75/76 → 77 working chain; DRAT ≤4 lock blocks.
  • m = 8 (8-gons)N(8,6,6) ≤ 61115-row CNF+DRAT replayed; signotope N(8,8,8) ≤ 318 (planar ES(8) ≤ 319 is Schroeder’s bound).
  • m = 9 (9-gons)ES(9) ≤ 1148Native-trace duals C₉ ≤ 688, D₉ ≤ 458 replayed → working 1148 (vs. 1508 / 1501*).

Why it matters Replayed CNF+DRAT and native-trace integer dual certificates tighten key signotope capacities and, via R7/R20 assembly, lower the working planar nonagon ceiling from 1508 (1501*) to 1148 alongside ES(7) ≤ 77.

Method Combine coordinate-slice double counting, shared-label CEGAR kernels and generative sparse integer duals.

Replayed duals/DRAT · assembly review pendingResults & methods ↗EN + 中文Earlier m = 7 ↗中文
Scope & verification

Five finite suites are replayed directly in-repo: milestone9 N(8,6,6) ≤ 61 (115 shared-label rows + DRAT; the 9-hole hitting vector is a matrix witness, not a 61-point signotope), zero-float native-trace integer duals N(9,8,8) ≤ 688 and junction-restricted D₉ ≤ 458 (matching the 47,635-row depth-≤3 LP ceiling), Schroeder’s 15-file octagon omission trees (193 / 124), DRAT exclusion of 5 common m = 7 lock blocks, and the 25-point E21 abstract witness (with historical same-family/provisional acceptance of N(7,5,6) ≤ 25). Planar ES(8) ≤ 319 is Schroeder’s published preprint bound; all structural, cyclic-sign transfer (N(8,8,8) ≤ 318) and R7/R20 assembly manuscripts (ES(7) ≤ 77, ES(9) ≤ 1148) remain working proofs pending independent mathematical review, with no Lean formalization.

How a problem moves through MathRSI

Parallel agents develop proofs. Cross-model reviewers challenge them. New evidence returns to the right stage and becomes reusable research memory.

  1. Problem
  2. Recognize the domain
  3. Load its plugin, or the shared base alone
  4. Adaptive stages

Research wikiread before each obligation · write lemmas, counterexamples and failures · reuse in the next step

Does this reduction make the original problem easier?Global progress checks · v4 maintenance on main

At planned checkpoints—or when new local results leave the same bottleneck untouched—two independent agents reassess the whole problem. One checks what remains even if the current work succeeds; the other compares alternative routes and the total cost of finishing.

The coordinator then changes the next task: connect a missing proof step, change the representation, switch routes, assemble a sufficient result, or continue within a concrete budget. Necessary complete calculations retain their completion plan.

Read the execution protocol ↗

Research replay · Erdős 86 · shown in the current workflow