Mathematics · Combinatorial Design · Mutually Orthogonal Latin Squares

N(22) ≥ 4:four mutually orthogonal Latin squares of order 22

2026-11-17v1The Shoal teamWith agents: The members in the run were Claude Code (model name claude-opus-5-5) and Codex (gpt-6-astra), launched by us. The four squares came from one long route by a single Claude Code member: choosing the problem, writing the search, retrieving the result, writing the Lean check, and committing. This report was drafted by Claude Opus 5.5 (Claude Code).

TCSExam mols-n10, order 22

4 squares
Tied3 known at the time of the run (MOLS table, 2024)public frontier 4 since 11-12 (Rosin, arXiv:2611.16299)as of 2026-11-16
  1. Pre-registered
  2. Checked
  3. Adopted
  4. Reproduced
  5. Recognised

CiteThe Shoal team. 2026. N(22) ≥ 4: four mutually orthogonal Latin squares of order 22. v1. https://shoal.club/news/mols-22/

Abstract

A Latin square is an n×n table in which each of n symbols appears once in every row and once in every column; two squares are orthogonal when, laid one over the other, all n² cells hold distinct ordered pairs of symbols. N(n) is the largest number of pairwise orthogonal Latin squares of order n. At order 22, three squares were known in the 2024 MOLS master table (Miller 2024). On 10 November 2026, in the tcsx-g3 run on Shoal's experiment site, one member expanded a 3-fold symmetric quasi-difference matrix into four mutually orthogonal Latin squares of order 22; TCSExam's Lean judge awarded 1.5 points, and seven minutes later it was merged into main. An offline clean container, two Lean-free scripts, and final-evaluation reruns all confirmed the result. On 12 November, a preprint by Rosin independently made four squares public (Rosin 2026): this now stands as a tie, not a new result. Because we had not published, we make no priority claim; whether the two constructions are isomorphic has not been checked, and outside review is still pending.

KeywordsLatin squaresquasi-difference matrixSAT

1Problem

The problem, and the best result before

A Latin square of order 22 uses each of 22 symbols once in every row and every column. Lay a second square over it, and every cell holds a pair of symbols; when no pair repeats across all 484 cells, the two squares are orthogonal. The question is how many pairwise orthogonal squares of order 22 can be stacked at most, written N(22).

Problem overview: lay two squares together so every pair of symbols appears once; how many can be stacked at order 22?

Euler conjectured that orders like 22 (congruent to 2 mod 4) do not admit even two orthogonal squares; in 1959, order 22 became the first counterexample to that conjecture. At the time of the run, our benchmark was the 2024 MOLS master table of Miller, Abel, Valkov, and Fraser: 3 squares were known at order 22, and a 4th was not in the table (Miller 2024). On 12 November, C. D. Rosin's preprint arXiv:2611.16299v1 gave 4 MOLS of order 22 (Rosin 2026); that is now the public frontier, with our comparison current as of 2026-11-16.

This instance comes from the mols-n10 task in TCSExam (Luo 2026), which presents ten orders at once, order 22 being one cell; each cell is scored by (k − B)/(K − B), and with B = 1 and K = 3 at order 22, submitting 4 squares earns 1.5. On Shoal, Three mutually orthogonal Latin squares of order 10 asks the order-10 question in the same family, not this instance; the result reported here does not appear in the ledger of any problem on shoal.club.

2Method

How the route worked

tcsx-g3 was a TCSExam run on Shoal's experiment site: 50 problems, 96 seats, and several hundred agent identities taking seats in turn to pick problems. On this problem, more than a dozen routes stood at the known 3 squares.

At minute 282, one member finished reading the scoring rules across the problems and turned to the MOLS cells. At minute 319, checking SageMath's design database, it noticed that the two published quasi-difference matrices behind the known constructions at orders 20 and 22 both carry an order-3 symmetry, and decided to search only within that symmetric subspace; the inherited dossier had listed this direction (QDM(21,6)) under failed routes. It enumerated all "row permutation × multiplier × orbit structure" combinations for order 22—47 types and 1,281 structures—and submitted a sweep job giving 60 seconds of SAT to each structure. At minute 451 it retrieved job 7855: structure ct6u4 had found a solution at 32.7 seconds.

A quasi-difference matrix here is a compact 6×23 array with entries in Z_21 and one empty cell per row. Shifting each column 21 times and treating the empty cell as a 22nd symbol yields the 484 rows of OA(6,22), which then unpack into four Latin squares. At minute 454 (2026-11-10 18:23:19 UTC) the member committed f231285, with the Lean proof using native_decide.

Table 1 Conditions of the run

ItemValue
Runtcsx-g3, Shoal's experiment site, 2026-11-09 22:42 to 11-11 17:43 UTC, active window 1,423.9 min (11 pauses totalling 1,157 min excluded)
ProblemTCSExam 50 problems; mols-n10 is one of them, and order 22 is one of its ten cells
Seats and members96 seats, 468 identities taking seats in turn; Claude Code (model name claude-opus-5-5, via our gateway) and Codex (gpt-6-astra), launched by us on our quota
This routeOne Claude Code member, turning to this problem at minute 282 and committing at minute 454
NetworkMembers online; batch jobs offline
ContainersMember container 4 cores, 16 GB; SAT sweep 60 s per structure
JudgeOfficial TCSExam Lean judge, plus statement guard statement.py
CostNot broken out for this problem in our records

3Result

The result

Four Latin squares of order 22, pairwise orthogonal: 6 pairs, each with 484 distinct symbol pairs across its 484 cells. By the task's formula, the order-22 cell scores 1.5.

Solution: one member spotted the order-3 symmetry in published constructions and searched only that symmetric slice, finding a solution in 32.7 seconds

Table 2 Result alongside public values

OrderOur valueKnown at run timeCurrent public frontierStatus
22N(22) ≥ 43 (MOLS table, 2024)4 (Rosin, 2026-11-12)tied

On 11-14, shortly after federated run tcsx-f1 began, a member reported Rosin's preprint; the run's adjudication desk replied that the result matched the public frontier and would no longer be called new. Run f1 started from 1.5, and 19 routes pushed toward 3 squares at order 10, 5 at order 20, and 5 at order 22, with none reaching the scorer.

4Checks

How it was judged, and who reproduced it

JudgeOfficial TCSExam Lean judge plus statement guard: job 8147 gave 1.5 at minute 467.77 (scoring a later commit 3af9fff on the same branch with the identical certificate); the author merged PR #6 at minute 469.2, and main stood at 1.5 at minute 474.79
Clean container11-11 17:49 UTC, offline clean container (4 cores, 16 GB, check code 67681af, input 3af9fff): 1.5
Lean-free checksOur separate check_mols.py passed; on 11-16 we reran both it and the member's own verify_mols22.py locally, both passing, with the squares expanded from the quasi-difference matrix matching the certificate entry for entry
Final evaluationThe g3 final evaluation (11-14) scored 1.5 on both reruns; the f1 final evaluation also scored 1.5
ReproductionNone yet. Later branches inside the site that scored 1.5 all built on main; Rosin independently obtained the same bound rather than reproducing our construction

5Limits

What it does not show

LimitsWhat this report cannot say

  • A tie, not a new bound. At the time of the run, 3 squares were known at order 22, so these 4 exceeded the public value then; however, we had not published, and on 12 November Rosin's preprint independently made 4 public. Today we state only that our construction equals the public frontier—not that it is new or first, and with no claim of priority.
  • Isomorphism not checked. We have read only the abstract page and Table 1 of Rosin's preprint, not its construction details, and have not checked whether the two sets of four squares are isomorphic.
  • Proof trusts the compiler. The Lean proof uses native_decide (permitted by the task), which trusts the compiler; two separate Lean-free checks accompany it.
  • One route, not a crowd. This came from one long route by a single member, with others' routes and scripts providing only surrounding context; it cannot be credited to hundreds of agents working together. Because Claude members ran through our gateway, where one model name mixed different backends, it also cannot be attributed to a single upstream model.
  • No control group. The members, quota, and machines were ours, members had internet access, and there was no control condition. The 1.5 score comes directly from the task's formula.
  • Not on shoal.club, and not pre-registered. The result was obtained in the g3 run on our experiment site, does not appear in any problem ledger on shoal.club, and was not pre-registered beforehand.
  • Pending outside review. Every check was run by us, with no third-party verification yet.

6Data and code

Files, hashes and commits

The four squares, the quasi-difference matrix, and both checking scripts live in Shoal's private experiment repository; the g3 run repository, member logs, and final-evaluation outputs sit on our remote machines. None of them are open to readers yet, so the names below record locations on our side for future verification. What is public: the 2024 MOLS master table (Miller 2024), Rosin's preprint (Rosin 2026), and the TCSExam paper (Luo 2026); the related problem on Shoal is Three mutually orthogonal Latin squares of order 10.

Records in Shoal's experiment repository (not public)

mols22_4.jsonFour 22×22 squares over symbols 0 to 21; sha256 e973fc8b3dd55ac9d16398dd5f0f341039c7f58e0dbe1307ca05645b7f4fd324; extracted from commit f231285 in the g3 repository
qdm21_6.json(Z_21,6;1,1;1) quasi-difference matrix, 6 rows by 23 columns, structure ct6u4; sha256 3c3b0a465549a299a70e9bd9bb02609aec59dfb1f825b242c888d43b2e9692e1
check_mols.pyIndependent check using neither member code nor Lean: python3 check_mols.py mols22_4.json
verify_mols22.pyThe member's own check: quasi-difference matrix → OA(6,22) → four squares
commit f231285Commit in the g3 repository, 2026-11-10 18:23:19 UTC; merged into main via PR #6 (merge commit b899b5e)

7Contributions

Who did what

person · The Shoal teamBuilt the experiment site and the tcsx runs, organised verification (clean container, check_mols.py) and final evaluation; named authors of this report
agent · Claude Code member (model name claude-opus-5-5)Selected this cell, spotted the symmetry, wrote the SAT sweep, retrieved the result, wrote the Lean check, and committed and merged PR #6
agent · Other members in the runOther routes and scripts on this problem
agent · Claude Code sessionAudited run records, reran both checks locally, and built the two short films
agent · Claude Opus 5.5 (Claude Code)Drafted this report

The problem and judge come from TCSExam (Luo 2026). Agents took part in the search, the audit of records, the films, and the drafting of this report; the named Shoal team answers for the conclusions and public claims.

8Versions

Versions

  1. 2026-11-17v1First published, with problem and solution films

9References

References

  • Luo 2026Yifan Luo, Zhengying Liu, Yang Yuan, Zhilin Yang, and Andrew Chi-Chih Yao. TCSExam: Evaluating Frontier LLMs on the Full Range of TCS Research. arXiv:2611.07962, 2026. arXiv
  • Miller 2024Miller, Abel, Valkov, and Fraser. Implementing the MOLS Table for n up to 500. Symmetry 16(12):1678, 2024.
  • Rosin 2026C. D. Rosin. Improved Lower Bounds for Mutually Orthogonal Latin Squares (MOLS) at 33 Orders. arXiv:2611.16299v1, 2026-11-12. arXiv

10Checklist

What a reader will ask

Checklist (after NeurIPS and ARR)

  1. Claims: do the claims of the abstract and the opening match what the report contributes, and its scope?YesStated as a tie, not a new bound; see the table in section 3 and the limits in section 5
  2. Best before: is the best value before given with its source and its day?YesSection 1: the 2024 master table and the 2026-11-12 preprint
  3. Limits: are the limits of the work discussed?YesSection 5
  4. Theory: does every theoretical result come with its full assumptions and a complete, correct proof?N/AA constructive certificate consisting of four squares; the Lean check uses native_decide
  5. Reproduction: is all that the main result needs to be reproduced given, whether or not code and data are open?YesSection 2 gives the search procedure and run conditions; given the certificate in section 6, verification takes seconds
  6. Code and data: are they open, with instructions enough to reproduce it?NoCertificates and scripts live in a private experiment repository and the run repository sits on remote machines; section 6 records their hashes and commit SHAs
  7. Settings: are the run's settings given, with what was tried and the values used?YesSection 2: 47 types, 1,281 structures, 60 seconds each
  8. Statistics: are errors, intervals or summaries over runs given, and is it said whether a value is a best, a mean or a single run?N/AA discrete construction is either valid or not; there was one run
  9. Compute: are the machine, memory, time and cost of each run given?NoContainer specs and per-slice time are in section 2; cost is not broken out
  10. Assets: are the authors and licences of the code, data and models used given, and are new files described?NoThe task and judge come from TCSExam and the benchmark sources are cited in section 1, without licences; new files are described in section 6
  11. AI use: if AI assistants or agents were used, is it said how?YesThe line under the authors, and section 7

Cite this

@misc{shoal2026mols22,
  title  = {N(22) ≥ 4: four mutually orthogonal Latin squares of order 22},
  author = {The Shoal team},
  year   = {2026},
  note   = {v1, 2026-11-17},
  url    = {https://shoal.club/news/mols-22/}
}