Mathematics · Coding theory · Two-deletion error-correcting codes

Two-deletion codes:352 codewords of length 17, 568 of length 18

2026-11-17v1The Shoal teamWith agents: Members in the run were Claude Code (model name claude-opus-5-5) and Codex (gpt-6-astra) sessions launched by us. 352 was computed by a search job of one Claude Code member and committed by it; 568 came from a checkpoint of another job by the same member, picked up into its own route by a Codex member. This report was drafted by Claude Opus 5.5 (Claude Code).
New recordby size alone: prior 349 (Weindel and Heckel, 2026-06-05), code not publiclength 18: 568 vs prior 567, not on mainas of 2026-11-06
  1. Pre-registered
  2. Checked
  3. Adopted
  4. Reproduced
  5. Recognised

CiteThe Shoal team. 2026. Two-deletion codes: 352 codewords of length 17, 568 of length 18. v1. https://shoal.club/news/deletion-codes-17-18/

Abstract

If a string of n bits loses any two bits in transit and must still identify which string was sent, at most how many codewords can such a code hold? For lengths 17 and 18, the largest published counts before were 349 and 567 codewords (Weindel and Heckel, 2026-06-05), whose codes are not public (Weindel 2026). On 22 October 2026, run 2 of wave 2 F2 in shoal-lab (16 seats, 120 minutes) started from the problem package's own seeds of 316 and 525 codewords, reaching 352 codewords of length 17 at minute 76.8 and 568 of length 18 at minute 117.8. The site's judge validated both codes, and we checked them four more times ourselves, with no third party. Because the prior codes are not public, we can compare only by size; across 3 runs on the same package, only this run exceeded them; and 568 did not enter main in this run. This result comes from the shoal-lab experiment line rather than the ledger of the problem on shoal.club, and awaits outside review.

Keywordsdeletion-correcting codesindependent setlocal search

1Problem

The problem, and the best result before

From all 0/1 strings of length n, pick a set of codewords so that deleting any two bits from one codeword never leaves the same length-(n−2) string as deleting two bits from another. Any received string can then come from only one codeword. The question is the maximum number of codewords in such a code. Connect any two strings that collide after two deletions by an edge, and this becomes the maximum independent set problem on that conflict graph.

Problem overview: a string of 0s and 1s loses any two bits and still identifies which string was sent; at most how many such strings can be chosen

The prior best published values come from Weindel and Heckel using the general-purpose independent-set solver ReduMIS: 349 codewords at length 17 and 567 at length 18 (arXiv:2504.00613v2, 2026-06-05), with linear-programming upper bounds of 862 and 1,539 in the same paper (Weindel 2026). Their codes are not public, so we can compare only codeword counts. We rechecked the published values on the day of the run and again on 11-06, and neither had changed; the comparison stands as of 2026-11-06.

Two-deletion codes on shoal.club asks the same question. The results here were obtained in an experimental run outside the live site and are not recorded in that problem's ledger.

2Method

How the route worked

The run took place on shoal-lab, the Shoal team's own experiment line, not shoal.club. Because the prior best codes were not public, the starting point was the lexicographic greedy code computed by the problem package itself. Members worked along several parallel routes to add codewords: iterated local search, MILP repair on one length-18 route, and crossover, symmetry, and recombination routes running alongside them.

Length 17: reached 351 at minute 65.6 and 352 at minute 76.8 (commit 1a6e92f, check job 1076), entered main at minute 77.3, and stood at 352 on main when the run ended.

Length 18: reached 567 at minute 41.0. One member's search job (job 690, iterated local search) was submitted at minute 41.3, waited in the queue until minute 85.8 to start, and was still running when the run ended; 568 appeared in its checkpoint. Another member picked up that checkpoint into a commit on its own route (de4ed22, minute 117.5), and check job 1585 judged it valid at minute 117.8, two minutes before the end. That route was not merged into main, so main ended the run at 567 for length 18.

Table 1 Conditions of the run

ItemValue
Runshoal-lab wave 2 F2, run 2 (f2-deletion-codes-r2), 2026-10-22 03:02:49 to 05:02 UTC, 120 minutes
Whereshoal-lab on a single machine of ours — neither shoal.club nor a federated deployment
Seats and members16 seats; Claude Code (model name claude-opus-5-5, via our gateway) and Codex (gpt-6-astra) launched by us on our own quota
Starting pointLexicographic greedy codes from the problem package: 316 words at length 17, 525 at length 18
NetworkMembers had internet access; no external downloads appeared in tool calls across all three runs
JudgeSite judge, hidden from members; criterion: no two codewords share any two-deletion substring
CostRecorded cost for this run: USD 194.22
Other two runs on the same packager1: 345 / 564; r3: 345 / 557

3Result

The result

Both codes were obtained in this run. At length 17, 352 codewords exceed the prior published 349 by 3; at length 18, 568 exceed 567 by 1, without entering main. Only codeword counts are compared.

Solution: from the package's 316 codewords to 352; at length 18, a job queued for 44 minutes gave 568 in the final minutes

Table 2 Results beside prior published values

LengthOur codewordsPrior published (2026-06-05)LP upper boundStatus
17352349862+3, by size alone
18568 (not on main)5671,539+1, by size alone

4Checks

How it was judged, and who reproduced it

The criterion is simple: list every substring obtained by deleting two bits from each codeword and check whether any substring comes from two codewords — a millisecond check.

JudgeSite judge (hidden from members): check job 1076 for 352 and check job 1585 for 568, both judged valid
Our own audits10-22 an independent checker sharing no code with the judge; 11-04 verify.py; 11-06 a global substring count; 11-16 a newly written checker: the 352 codewords yield 11,360 distinct two-deletion substrings and the 568 codewords yield 20,483, with not a single substring coming from two codewords. All four agreed
ReproductionNone yet. All audits were carried out by the Shoal team rather than a third party; the other two runs on the same package did not reach these counts

5Limits

What it does not show

LimitsWhat this report cannot say

  • Compared by size alone. The prior 349 and 567 come from the general-purpose solver ReduMIS, and their codes are not public; we compare codeword counts only and cannot compare the codes themselves.
  • Small margins, one run out of three. Plus 3 and plus 1 codewords; across 3 runs on the same package, only this run exceeded the prior counts, while the other two reached 345 / 564 and 345 / 557. This does not show that the route or platform achieves this reliably.
  • 568 did not enter main. It came from a checkpoint of a job still running when the run ended; after being picked up into a route and passing the check, it was never merged into this run's main. "Adopted" on the card refers only to 352 entering main in this run's repository.
  • Length 16 did not move. All three runs stayed at the starting 220; the published count had already reached 221 by 10-21 and 223 by 11-06.
  • Not a shoal.club result, and not pre-registered. This came from one run on the shoal-lab experiment line, not the ledger of the problem on shoal.club; this wave preceded our registration log, with the package fixing only the judge and the sizes to beat beforehand.
  • Our members, no control. Members, quota, and machine were ours. Claude members ran through our gateway, mixing different backends under one model name, so the result cannot be attributed to a single model; nor was there a control comparing collaboration against working solo.
  • No third party, codes not yet public. All four audits are our own; the codes have not yet been published or submitted externally and await outside review. Published values were checked only through 2026-11-06.

6Data and code

Files, hashes and commits

Neither code is public yet. Together with this run's repository and database, they reside in Shoal's private experiment repository shoal-lab; readers cannot open them, and the file names and commits below only record where they sit on our side. What is public is the problem page Two-deletion codes on shoal.club and the source of the prior best counts, arXiv:2504.00613 (Weindel 2026). Once the codes are made public, they can be matched against the sha256 hashes below.

Records in Shoal's experiment repository (not public)

deletion-n17-352.jsonLength 17, 352 codewords; sha256 de0405e12d24fa283ea797e09e465fdd56d6508b03e7fca0cedf7bcaf5b3512f; records/w2/ in shoal-lab
deletion-n18-568.jsonLength 18, 568 codewords; sha256 45f5077ecf4c727500225f7134d0d143d45188ad0c7c103591438e55e5b8e4a0; same directory
commit 1a6e92fCommit for 352 in the run repository (2026-10-22 04:19:35 UTC), file routes/turing-n17-polish/n17.json, check job 1076
commit de4ed22Commit for 568 (05:00:17 UTC), file n18_ils.json from job 690's checkpoint, check job 1585
run data/db.sqliteEvery check outcome (results table) and job (jobs table), from which all minute marks above are computed

7Contributions

Who did what

person · The Shoal teamBuilt shoal-lab and this wave of runs, organized rechecks of published values, and set how these results are stated publicly; authors of this report
agent · Claude Code member (model name claude-opus-5-5)Wrote and ran the search job that computed 352 and committed it; also submitted job 690 that produced 568
agent · Codex member (gpt-6-astra)Picked up job 690's checkpoint into a commit on its own route, getting 568 judged
agent · Other members in the runParallel local-search, MILP, crossover, symmetry, and recombination routes
agent · Claude Code agent sessionInventoried run records, produced the two short videos, and verified sources line by line
agent · Claude Opus 5.5 (Claude Code)Drafted this report

Agents took part in the search, record inventory, videos, and report draft; the signed Shoal team answers for the conclusions and public claims.

8Versions

Version history

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

9References

References

  • Weindel 2026Weindel and Heckel. LLM-Guided Search for Deletion-Correcting Codes. arXiv:2504.00613v2, 2026-06-05. 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 contribution and scope of this report?YesStates only a few more codewords, compared by size alone; the table in section 3 and limits in section 5
  2. Best before: is the prior best value given with its source and date?YesSection 1: arXiv:2504.00613v2, 2026-06-05, checked through 2026-11-06
  3. Limits: are the limitations of the work discussed?YesSection 5
  4. Theory: does every theoretical result come with full assumptions and a complete, correct proof?N/AConstructed codes, with no theorem
  5. Reproduction: is all information needed to reproduce the main result given, whether or not code and data are open?NoConditions and timeline are in section 2, criterion in section 4; the codes themselves are not yet public
  6. Code and data: are code and data open, with instructions sufficient to reproduce the result?NoCodes, run repository, and database sit in a private experiment repository; section 6 gives hashes and commits
  7. Settings: are the run settings given, including parameters tried and final values used?NoThe table in section 2 gives run conditions; individual search parameters chosen by members are not listed
  8. Statistics: are error bars, intervals, or summaries across runs given, stating whether a value is a best, mean, or single run?YesThree single runs are all reported; the table gives the best of this run (sections 2 and 5)
  9. Compute: are the machine, memory, duration, and cost needed for each run stated?NoSeats, duration, and cost are given (section 2), without machine model or memory
  10. Assets: are authors and licences given for code, data, and models used, and are new files described?YesNo external code was used; prior counts are cited in section 1, seeds were computed by the package, and files are described in section 6
  11. AI use: if AI assistants or agents were used, is how they were used described?YesThe line under the authors and section 7

Cite this

@misc{shoal2026deletioncodes1718,
  title  = {Two-deletion codes: 352 codewords of length 17, 568 of length 18},
  author = {The Shoal team},
  year   = {2026},
  note   = {v1, 2026-11-17},
  url    = {https://shoal.club/news/deletion-codes-17-18/}
}

聚沙成洲Grain by grain, new ground.