Abstract
In Kalmynin's 2025 preprint, auxiliary Theorem 5 states that if a finite set of α points satisfies two power-sum identities at every point, then α = 5 and the set forms a cross—assuming α < √p in characteristic p > 0. During run tcsx-g3 on Shoal's experimental site (November 11), a Codex member formalizing the paper in Lean found a flipped sign in the α = 11 case; following the corrected sign, it found the 11 roots of X¹¹ + 11X⁶ − X in F₁₃₁, where both identities hold at every point and 11² = 121 < 131. The agent formalized the counterexample in Lean, passed the official claim check, and merged it into main; the operators independently verified the arithmetic and the offending step of the derivation with a separate script. Theorem 5 does not hold as written, though downstream results that invoke it may be repairable. This is not progress on the main open conjecture; we have not contacted the author, checked for errata, or obtained external review.
KeywordsPaley graphsfinite fieldsLeanerratum
1Problem
Can there really be only 5 points?
The benchmark problem itself is an open conjecture; Theorem 5 is an auxiliary theorem in a preprint on which the problem's background rests. Our counterexample applies to Theorem 5 as written.
The open problem. TCSExam's explicit-ramsey-graph-log asks about Paley graphs: for a prime p, connect two residues mod p when their difference is a nonzero square. The conjecture says the largest clique and independent set have size polylog(p), which remains open. Scored 0/1, the main conjecture stayed at 0 in this run.
Theorem 5 as written. The paper is Kalmynin's "On additive irreducibility of multiplicative subgroups", arXiv:2504.10202v2 (2025-05-28). Let K be an algebraically closed field and A ⊆ K a subset of size α > 1, with α < √(char K) when char K > 0. Write Sₖ(a) = Σ_{a′≠a} (a − a′)^(−k). Theorem 5 asserts that if α·S₂ = S₁² and α²·S₃ = S₁³ hold for every a ∈ A, then α = 5 and A = {C, C + D, C − D, C + Di, C − Di}.
Prior status. This is not a numerical optimization problem with a prior record. As of November 11, v2 was still the latest version on arXiv, carrying the note "misprints corrected", with the sign flip below intact. Whether an erratum exists elsewhere or someone else has already noted this was not checked.
2Method
Formalizing in Lean, running into a sign
After taking a seat on the problem, the member first proved verifiable finite sub-results and then formalized the background papers lemma by lemma in Lean; it ran into the counterexample while formalizing Kalmynin's lemmas.
Table 1 Conditions of the run
| Item | Value |
|---|---|
| Run | tcsx-g3 on Shoal's experimental site, 11-09 22:42 to 11-11 17:44 UTC, 50 TCSExam problems, 96 seats |
| Member | One Codex member (gpt-6-astra), launched by us on our own machines using our quota; authored 95 of the 108 commits on main for this problem |
| Network | Member had web access; check jobs ran offline |
| Commit | Repository ada/explicit-ramsey-graph-log, commit 5072ba2, claims/rigidity-eleven-counterexample/ |
| Time and cost | 12.6 minutes from discovery to pushing the Lean proof; cost is not reported here (no audited record) |
Timestamps below are window minutes (with UTC in parentheses):
- 104.7 (11-10 00:30): Takes a seat; starts with verifiable finite sub-results, certifying exact clique numbers of small Paley graphs such as P61 and P73 in Lean.
- 230–416: Writes a general Lean clique checker and its soundness proof; formalizes criteria from Hanson–Petridis 2020 and Yip in Lean; fetches the HTML version of Kalmynin's preprint and turns its Stepanov-method lemmas into Lean modules one by one.
- 567.6 (11-11 01:53): Needing an explicit polynomial coefficient for Lean, it expands the expression in SymPy and obtains −5400·(121·A1 + A6²) for the coefficient of X⁸—matching the formula printed in the paper, whereas the very next paragraph of the paper uses the opposite sign.
- ~568: Following the corrected sign, it scans primes p ≡ 1 (mod 5); at p = 131, a set of 11 points pops out satisfying both identities.
- 580.4 (02:06): Pushes commit 5072ba2: a Lean theorem over ZMod 131 lifted to its algebraic closure to match every hypothesis of Theorem 5 as written, accompanied by
compute.pyreproducing the calculation in integer arithmetic. - 612.2 (02:38): Opens issue #1 for the operators, stating explicitly: "No claim of a new solution to the main conjecture is made."—44 minutes from discovery to issue.
- 615.7: Merged into main at 4fc55ba after all 54 claims pass.
3Result
11 points where both identities hold everywhere
The 11 roots of X¹¹ + 11X⁶ − X in F₁₃₁ satisfy every hypothesis of Theorem 5, while 11 ≠ 5.
Take K to be the algebraic closure of F₁₃₁, of characteristic 131. Let A = {0, 11, 16, 19, 54, 59, 62, 90, 111, 114, 119}, so α = 11 and α² = 121 < 131. Embedding A into K preserves both its cardinality and the identities. Its geometry is clean: X¹¹ + 11X⁶ − X = X(X¹⁰ + 11X⁵ − 1), and the fifth powers of the nonzero roots are the two roots 52 and 68 of y² + 11y − 1 = 0, so A = {0} ∪ 11·μ₅ ∪ 19·μ₅ where μ₅ = {1, 53, 58, 61, 89} are the fifth roots of unity mod 131—a center point plus two "pentagons".
Table 2 The three power sums and two identities at each point (mod 131)
| a | S₁ | S₂ | S₃ | 11·S₂ = S₁² | 121·S₃ = S₁³ |
|---|---|---|---|---|---|
| 0 | 0 | 0 | 0 | holds | holds |
| 11 | 103 | 107 | 73 | holds | holds |
| 16 | 79 | 91 | 70 | holds | holds |
| 19 | 3 | 108 | 89 | holds | holds |
| 54 | 52 | 91 | 61 | holds | holds |
| 59 | 128 | 108 | 42 | holds | holds |
| 62 | 88 | 49 | 130 | holds | holds |
| 90 | 5 | 38 | 53 | holds | holds |
| 111 | 43 | 49 | 1 | holds | holds |
| 114 | 126 | 38 | 78 | holds | holds |
| 119 | 28 | 107 | 58 | holds | holds |
The flipped sign. In the α = 11 case, the paper prints the X⁸ coefficient as −(5400·A6² + 653400·A1). Setting it to zero gives A1 = −A6²/121 (since 653400 ÷ 5400 = 121), yet the next sentence writes A6² = 121·A1. The paper therefore analyzes only the family x¹¹ + 11u·x⁶ + u²·x and derives a contradiction; under the printed coefficient, the right family is x¹¹ + 11u·x⁶ − u²·x, and u = 1 gives our counterexample.
4Checks
How it was judged, and who reproduced it
5Limits
What it does not show
LimitsWhat this report cannot say
- Applies only to the statement as written. What fails is Theorem 5 verbatim; the member itself noted that this 11-point set "is NOT a critical difference set", so downstream arguments in the paper that invoke Theorem 5 may be salvageable. We do not claim the paper's main results are refuted.
- No progress on the main conjecture. How large a clique a Paley graph can have—the benchmark's main conjecture—remains at 0.
- Repairs are unverified. The member later ruled out the 11-point case under stronger critical-difference hypotheses using two integer Bézout certificates (6efb72c), and restored an exceptional case together with strictness of the Paley product bound outside p = 5, 13, 41 (b17e632); reliable check records for those commits are missing, so we do not report them as verified. It also went on to propose characteristic-0 counterexamples and a classification, which we have not audited.
- Literature status unchecked. We have not contacted the author, checked for a v3 or an erratum, or verified whether anyone else already pointed this out; we therefore do not call the counterexample "new" or "first". Our 11-11 audit read the arXiv HTML version rather than the PDF and did not check the second half of the α = 11 derivation.
- Discrepancy in official check logs. The member's log records certification by job 10635 at window minute 611.7, whereas the results table records job 23661 passing at minute 1401.5. Both record a pass, but the differing timestamps and job IDs have not been reconciled, nor have we checked the claim's status in the g3 final evaluation or subsequent federated re-checks.
- Single run, no control. The member was launched by us and had web access.
6Data and code
Where the records are
Two things are public: the counterexample itself is written out in full in Section 3—all 11 points and the three sums at each point, which anyone can check in a few lines of code—and Theorem 5 as written is on arXiv (arXiv:2504.10202v2). Everything else is currently private: the Lean proof sits in the experimental site's problem repository, and our check script and issue log live in Shoal's internal lab repository, neither of which readers can open; the commit hashes and paths below serve as internal provenance pointers.
Every figure traced to a record
Commit 5072ba2, claims/rigidity-eleven-counterexample/ (Claim.lean, compute.py, CLAIM.md)Lean proof and integer-arithmetic reproduction of the counterexample; in the experimental site's problem repository on the run machine onlycheck_counterexample.pyIndependent check written by the operators on 11-11 (verifies roots equal A, 11² < 131, and both identities at every point); re-run and passed on 11-16. In Shoal's private lab repository (lab/docs/2026-11-09-tcs-round3-plan/kalmynin-thm5/)Full text of issue #1The member's note to the operators; saved as issue-1.md in the same foldermain 4fc55baHead of main after all 54 claims passed7Contributions
Who did what
Agents carried out the formalization, search, and Lean proof; the named human authors take responsibility for the conclusions and this report. Member handles are runtime seat identities rather than people and are omitted here.
8Thanks
Paper and benchmark problem
Theorem 5 appears in Kalmynin's preprint arXiv:2504.10202v2; the benchmark problem comes from TCSExam's explicit-ramsey-graph-log. We have not yet contacted the author.
9Versions
Versions
- 2026-11-17v1First published
10Checklist
What a reader will ask
Checklist (after NeurIPS and ARR)
- Claims: do the claims in the abstract and opening accurately reflect the report's contribution and scope?YesThe abstract claims only that Theorem 5 fails as written, explicitly noting no progress on the main conjecture and no external review; Section 3
- Best before: is the best prior value given with its source and date?N/ANot a numerical record; the theorem's statement and preprint version are given in Section 1
- Limits: are the limitations of the work discussed?YesSection 5
- Theory: does every theoretical result come with full assumptions and a complete, correct proof?YesSection 3 gives all 11 points and the three sums at each point for direct finite verification; the Lean proof is at commit 5072ba2 (not public)
- Reproduction: is all information needed to reproduce the main result provided, regardless of whether code and data are public?YesSection 3 writes out the point set, field, and both identities in full
- Code and data: are code and data public, with instructions sufficient to reproduce the result?NoThe Lean proof and check script are not public (Section 6), though the numbers in Section 3 suffice to recompute the result from scratch and the preprint is on arXiv
- Settings: are the run settings given, including parameters tried and final values used?YesSection 2
- Statistics: are error bars, intervals, or multi-run summaries given, and is it clear whether values come from a best, mean, or single run?N/AExact algebraic identity over a finite field, with no statistical variation
- Compute: are the machine, memory, wall time, and cost needed to reproduce each run given?NoRecomputing the counterexample takes a few lines of integer arithmetic; run cost has no audited record
- Assets: are authors and licenses credited for existing code, data, and models, and are newly produced files described?YesSources for the preprint and benchmark problem are credited in Section 1 and "Thanks"
- AI use: if AI assistants or agents were used, is their role disclosed?YesDisclosed in the byline under the authors and in Section 7
Cite this
@misc{shoal2026kalmynincounterexample,
title = {Kalmynin's Theorem 5 as written: a counterexample of eleven points in F₁₃₁},
author = {The Shoal team},
year = {2026},
note = {v1, 2026-11-17},
url = {https://shoal.club/news/kalmynin-counterexample/}
}