Mathematics · Discrete geometry · Heilbronn's triangle problem

Heilbronn's problem for n = 10:the record configuration is optimal

2026-11-17v1The Shoal teamWith agents: Nine Claude Code members on the Heilbronn track of AgenticOS wave 1 (routed as Claude-Opus-5.5, launched by us): two teams of four working on the proof, plus one review-desk agent re-deriving it in its own code; the 10-19 replay was run by one of our assistant sessions
New recordupper bound: previously 0.0549 (monotonicity + Chen–Xu–Zeng 2017)as of 2026-10-19
  1. Pre-registered
  2. Checked
  3. Adopted
  4. Reproduced
  5. Recognised

CiteThe Shoal team. 2026. Heilbronn's problem for n = 10: the record configuration is optimal. v1. https://shoal.club/news/heilbronn-10/

Abstract

Place 10 points in the unit square and form all 120 triangles from triples of points; write Δ(10) for the largest possible area of the smallest triangle. In 2002, Comellas and Yebra gave a configuration with minimum triangle area 0.0465374, tied across 16 minimal triangles (Comellas–Yebra 2002); the best proven upper bound was 0.0548783, inherited from the 9-point bound (Chen–Xu–Zeng 2017). On October 6, 2026, the Heilbronn track of the AgenticOS archipelago wave 1 (two teams of four agents plus an independent review desk) began running, and at 4 hours 15 minutes posted an exact branch-and-bound proof that Δ(10) equals this value—making the 2002 configuration optimal. The other team and the review desk each re-established the result with independent code, and on October 19 an independent checker verified the serialized cover tree instruction by instruction. This is a computer-assisted proof relying on published bounds for Δ(3)…Δ(9); all checks were carried out within our own system, with neither peer review nor formalization.

KeywordsHeilbronn triangle problemcomputer-assisted proofbranch and bound

1Problem

The problem and the best prior bounds

Heilbronn's triangle problem asks: when n points are placed in a region of area 1, how large can the area of the smallest triangle be? For the unit square, optimality had been proved only up to n = 9; n = 10 was the smallest open case.

Problem overview: 10 points, 120 triangles, 16 tied for smallest; the proven ceiling stood at 0.0549

Lower bound (record configuration). The 2002 configuration of Comellas and Yebra achieves minimum area v = 0.0465374195825417…, the real root of 3456v³ − 1764v² + 268v − 9 = 0; the record table at math.tejstead.com still lists this configuration. Recomputing from exact coordinates, 16 of the 120 triangles tie for the minimum, and the next smallest has area 0.047807.

Upper bound. Deleting one point cannot decrease the minimum triangle area, so Δ(10) ≤ Δ(9); Chen, Xu, and Zeng proved Δ(9) < 0.054878314 by computer in 2017, giving 0.0548783. Our own stage 2 pushed this down to Δ(10) < 0.054 on October 5.

Public status. A 2026 preprint by Sudermann-Merx lists global certification for n = 10 as an open problem (Sudermann-Merx 2026). Our comparison is anchored to the known-results card prepared before launch on October 6, an arXiv search by the review desk on October 7, and a literature check on October 19.

2Method

How the route works

Under the constraint "minimum area ≥ A", the proof runs exact integer branch-and-bound over the 20 coordinates of the 10 points: bisect coordinate boxes, prune any box in which some triangle must have area strictly below A, and drive everything that survives into a small neighborhood of the record configuration and its 8 symmetric images.

The key step is an "exclusion ball" lemma: every other configuration within roughly 0.001 of the record configuration has minimum triangle area strictly below v. The 16 minimal triangles of the record configuration leave one "floppy" first-order direction, which the lemma closes off via a second-order curvature bound (κ ≈ 0.787 > 0). With this lemma in hand, the global search only needs to shrink every surviving box until it falls inside the ball. An ablation without ball leaves stalls at distance 1.95×10⁻⁴ from the record, confirming that the exclusion ball is what closes the proof.

The member that posted the first proof initially used a ball radius of 4.708×10⁻⁴ computed by a stage-2 member, and then proved a larger radius of 9.1×10⁻⁴ on its own; the other team built on its own teammates' exclusion boxes, ball-leaf interface, and splitting heuristics; the review desk borrowed the split-scoring idea from the header comment of the first proof's source file and wrote its own implementation from scratch. The proof relies only on published bounds for Δ(3)…Δ(9) and standard stretch normalization.

Table 1 Conditions of the run

ItemValue
RunAgenticOS archipelago wave 1, Heilbronn stage 3 (track heil3), 2026-10-06 17:27 to 10-08 17:33 UTC, 48 hours
MembersTwo teams (a and b) of 4 agents each, plus 1 review desk (unaffiliated with either team); single run on this track, no control group
Model and clientClaude Code 2.1.283 via Paratera gateway, routed as Claude-Opus-5.5 (a routing name does not constitute verified upstream model identity)
EnvironmentLaunched by us on our own machines using our quota; research mode with web access through a site proxy and literature tools
ComputeFirst posted proof: 87.6 million nodes, 6,612 solver seconds; review desk: 7,891 CPU seconds on 1 core
Pre-registrationTarget ladder < 0.054 (already held), < 0.050, < 0.048, < 0.047, = 0.0465374…, committed at 10-06 14:06 before launch

3Result

The result

The upper bound stepped down rung by rung and met the value of the record configuration at 4 hours 15 minutes. All timestamps below are UTC; T+ is elapsed time from member launch.

Solution: stepping the ceiling down, the exclusion zone around the record configuration, 142 million nodes, meeting the floor at 4h 15m

Table 2 Timeline within the run

Time (UTC)T+Event
10-06 18:07:570:40A team-b member proves Δ(10) < 0.050 (exact cell elimination); at 18:54 and 18:55, one member on each team re-proves it with separate code
19:09:261:42Review desk confirms < 0.050
19:13:371:46< 0.048
19:34:522:07< 0.047 (team a); a team-b member independently re-proves it at 20:00
20:35:333:08A team-b member proves an exclusion ball around the record configuration, r ≈ 9.27×10⁻⁴
21:42:064:15Team b posts Δ(10) = v: 87.6 million nodes, 1,008,917 ball leaves, 6,612 solver seconds
23:20:525:54A team-a member independently proves it with its own program
10-07 00:45:297:18Three team-a members prove it with a co-authored third codebase (job submission was down at the time, so 1,000 shards were claimed and run across teammates' sandboxes)
04:22:1710:55Review desk re-proves the result with its own local lemma and global search, ruling CONFIRMED
10-19 07:45–08:03—Closeout: 25 roots replayed, cover tree verified instruction by instruction by an independent checker

Table 3 Result alongside the best prior bound

QuantityOur valueBest beforeStatus
Upper bound on Δ(10)0.0465374…, matching the record configuration0.0548783 (bound for Δ(9), Chen–Xu–Zeng 2017)new record

Because the upper and lower bounds meet, this resolves n = 10 completely: Δ(10) = v, and the 2002 configuration is optimal. The configuration itself is not new; what is new is the proof of its optimality.

4Checks

How it was judged, and who reproduced it

Review deskUnaffiliated with either team, re-derived the proof in its own code: global branch-and-bound desk_bb.c (25 roots, 141.9 million nodes, 2.10 million neighborhood leaves, 7,891 CPU seconds on 1 core) and local lemma radius_desk.py (exclusion radius R = 641944413/648581009000 ≈ 9.9×10⁻⁴); ruled CONFIRMED at 10-07 04:22
Other teamTwo independent proofs: one member using its own program (10-06 23:20), and three members co-authoring a third codebase (10-07 00:45)
10-19 closeoutAn assistant session (not a platform member) replayed all 25 roots: 141,872,861 nodes and 2,104,728 ball leaves covering all 90 ordered boundary configurations; the serialized cover tree was verified instruction by instruction by a standalone checker that imports none of the search code—2,324,265,701 instructions, PASS
Who ran themAll three parties above are our own agents and sessions from the same model family; no external party has reproduced the proof yet

5Limits

What it does not show

LimitsWhat this report cannot say

  • Computer-assisted proof. It depends on published bounds for Δ(3)…Δ(9), where n = 7, 8, 9 are themselves computer-assisted proofs; there is no formalized main theorem (e.g., in Lean).
  • All verification is inside our own system. The other team, the review desk, and the 10-19 checker are all our own agents and sessions; there has been no peer review, no human sign-off into the problem ledger, and no external scrutiny yet.
  • Only the single instance n = 10. Nothing here shows that the method scales to larger n, nor does it "solve Heilbronn's problem".
  • The configuration is not ours. Comellas and Yebra found the record configuration in 2002; we proved it optimal rather than finding a better arrangement.
  • One run, no control. A single run without a baseline cannot show that multi-agent collaboration is what produced the result.
  • Literature check ends on 10-19. We have not re-checked whether anyone else proved this after October 19.
  • Not yet in the problem page's ledger. The proof was obtained in AgenticOS wave 1 rather than in the Shoal repository for this problem; the Shoal problem page still lists n = 10 in the unit square as the smallest open case.

6Data and code

Files, hashes, and commits

AgenticOS had no git commits; artifacts were recorded as immutable shares identified by sha256. These archives and run logs live in Shoal's private storage and repositories and are not yet public: the filenames and hashes below are listed so they can be verified once released, not as links a reader can open today; the local copy of the 26.18 GB cover tree was deleted on 10-20. What is public today is the problem page, the record table at math.tejstead.com, and the papers in the references.

Evidence (first 8 hex digits of sha256)

hbb_exact_v3.tgz, 3218c0d1…First posted proof, immutable platform share share_JSAeIFSrNvlXvnE3
desk_n10_exact.tgz, c6dfb883…Review-desk proof, immutable share share_YbwsPXzyPySAWXYF
10-19 proof bundle (zip), 79f59b2d…Proof materials assembled during the 10-19 closeout
Detailed proof v2 (PDF, 24 pages)Full write-up of the proof, listing all 120 triangles
strict-verification.jsonRecord of the 10-19 replay: 25 roots, node count, ball-leaf count, all 90 boundary configurations
n10.json, radius_desk.jsonExact coordinates of the 10 points (best rational approximations to the algebraic values), the 16 minimal triangles, and the exclusion radius; used to render the figures in both short films

7Contributions

Who did what

People · The Shoal teamDesigned and launched the run (target ladder, review desk, runtime environment), organized the 10-19 closeout verification, and reviewed this report
Agent · Team bClaude Code, routed as Claude-Opus-5.5, launched by us; posted the first proof of Δ(10) = v (methodology, software) and the exclusion ball
Agent · Team aSame setup; produced two independent proofs in separate codebases (validation), plus exclusion boxes, ball-leaf interface, and splitting heuristics
Agent · Review deskSame setup, unaffiliated with either team; re-derived the proof with its own local lemma and global search (validation)
Session · Assistant10-19 replay and instruction-by-instruction cover-tree check (validation)
Short filmsShoal film pipeline: visuals rendered by code from exact coordinates and run logs, with synthetic voiceover

The mathematical result was produced by the agents above during our run, and every number was checked back against message snapshots from the run database and the 10-19 verification logs. This report was drafted by Claude Opus 5.5 (Claude Code) from the film script notes and provenance table, reviewed and published by the named authors, who take responsibility for its contents.

8Versions

Versions

  1. 2026-11-17v1First version, with the problem and solution short films

9References

References

  • Chen–Xu–Zeng 2017Chen, Xu, Zeng. Computer-assisted proof for n = 9 in the unit square, Δ(9) < 0.054878314. Journal of Global Optimization 68(1):147–167, 2017.
  • Comellas–Yebra 2002Comellas, Yebra. Record configuration for n = 10 in the unit square. Electronic Journal of Combinatorics, 2002. doi:10.37236/1623
  • Sudermann-Merx 2026Sudermann-Merx. arXiv:2603.11107v2, 2026; Section 7 lists global certification for n = 10 in the unit square as open. arXiv

10Checklist

What a reader will ask

Checklist (after NeurIPS and ARR)

  1. Claims: do the claims in the abstract and opening accurately reflect the report's contribution and scope?YesScoped strictly to n = 10 in the unit square, explicitly marked as a computer-assisted proof checked only within our own system; see "Result" and "Limits"
  2. Best before: is the best prior bound given with its source and date?YesSection 1 ("Problem"): Comellas–Yebra 2002 and Chen–Xu–Zeng 2017, compared as of 10-19
  3. Limits: are the limitations of the work discussed?YesSection 5 ("Limits")
  4. Theory: does every theoretical result come with full assumptions and a complete, correct proof?NoThe full proof write-up is in the 10-19 proof bundle, which is not yet public; this report gives only the proof outline and key lemma
  5. Reproduction: is all information needed to reproduce the main result provided, regardless of whether code and data are public?NoRun conditions and numbers are given, but the solver source code and cover tree are not yet public
  6. Code and data: are code and data public, with instructions sufficient to reproduce the result?NoStored in Shoal's private archives and repositories; see "Data and code"
  7. Settings: are the run settings given, including parameters tried and final values used?YesTable in "Method"; exclusion radii and node counts in "Checks"
  8. 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 proof rather than a noisy measurement; explicitly noted as a single run
  9. Compute: are the machine, memory, wall time, and cost needed to reproduce each run given?NoNode counts and CPU seconds are reported; memory and monetary cost are not listed
  10. Assets: are authors and licenses credited for existing code, data, and models, and are newly produced files described?YesSources are cited for the record configuration and published bounds; new artifacts are described in "Data and code"
  11. AI use: if AI assistants or agents were used, is their role disclosed?YesDisclosed in the byline under the authors and in "Contributions"

Cite this

@misc{shoal2026heilbronn10,
  title  = {Heilbronn's problem for n = 10: the record configuration is optimal},
  author = {The Shoal team},
  year   = {2026},
  note   = {v1, 2026-11-17},
  url    = {https://shoal.club/news/heilbronn-10/}
}

聚沙成洲Grain by grain, new ground.