← All posts
Competition runSquares in Squares · Kaggle Playground · $50,000 · closes 18 Jul 2027

Squares in Squares on Kaggle: 46 Rounds, 22.2502 to 22.1756, and a Moving Field

Our Squares in Squares run: 46 rounds took the public score from 22.2502 to 22.1756, rank 326 of 2,052. Which merges held and which of our ideas did not.

326 / 2,052
Public rank (14 Nov)
22.1756
Best public (lower wins)
22.1580
Best local
46 · 17 improved
Rounds

Manas Joshi·November 14, 2026·7 min read

Squares in Squares: competition facts
  • Where: Kaggle Playground, $50,000 prize pool, no Kaggle medals
  • Task: Pack n unit squares into the smallest square, for every n from 1 to 1,000
  • Score: Sum of L/√n − 1 over all 1,000 puzzles · lower is better
  • Rules: No two squares may overlap, checked in strict floating point
  • Dates: final deadline 18 Jul 2027
  • Field: 2,052 teams on the public leaderboard (14 Nov); leader 22.0500
  • More: all Squares in Squares facts, scoring and FAQ
🗓️
Standing on 14 Nov 2026: rank 326 of 2,052 (top 15.9%) with a best public score of 22.1756. Top 10% that day was 22.1617. Our newest local packing measures 22.1580, which would sit inside it, but it is not submitted yet. Playground competitions award no medals; the prize pool still applies.

The short version

Kaggle Overnight started on 5 November and has recorded 46 rounds, 40 of them launched unattended by auto-run. 17 improved the score, and 16 of those were certified merges of public packings: for each n, take the smallest packing anyone has published, check it for overlap with a safety margin, and keep it. Our own search ideas ran 12 times for one gain of 0.0000015. We improved the score almost every day and still fell from rank 174 to 326, because about 400 more teams joined and the whole leaderboard tightened.

Score timeline

DateRoundScoreWhat changed
5 NovR322.2502Margin-safe certified merge of 12 public outputs — submitted
6 NovR622.2367Newest public outputs, re-merged per n
7 NovR1022.2238Two more public portfolios merged
9 NovR1222.2173A public certified merge folded in
10 NovR2222.2004Merged on this machine after a Kaggle kernel failed — submitted
12 NovR2922.1926A validated public merge plus its source outputs — submitted
12 NovR3122.1793Another public merge — submitted
13 NovR4022.1756Every notebook the watcher had recorded — best public score
14 NovR4522.1580Notebooks claiming a better score, merged per n — local only

Squares in Squares is scored on the submitted CSV alone, so a local score is computed the same way as the public one, provided the packing passes Kaggle’s overlap check.

localhost:7842
Squares in Squares Runs page, each round launched by auto-run with Claude Code
Runs: every recent round says Claude Code · auto — launched by auto-run with no chat open
localhost:7842
Squares in Squares experiment tree with failed early merges and a dashed merge candidate
The first rounds as a tree: R1–R2 failed the full check, R3 held, R3 + R5 a merge candidate

What did not work

IdeaWhat happened
First per-n merges (R1–R2)Passed the quick screen, failed the full overlap check; fixed with a safety margin
Free-angle polish of every puzzle (R4)Hit the 30-minute round limit; the rerun finished with no gain
Neighbour transfer: n+1 packing minus one square (R7, R9, R25)No gain in three tries; one was interrupted
Tilted-block and corner-band seeds (R11, R27)No gain
Stretch smaller packings and insert a square (R28)No gain
Large remove-and-reinsert moves (R34)Crashed in a worker process; recorded, not retried blind
A merge submitted on 9 NovPassed our check, came back from Kaggle’s scorer as an error

The lessons

When everyone shares packings, the merge is table stakes. The field improves daily, so what mattered most was speed: the loop now watches public notebooks and merges new outputs as they appear.

localhost:7842
Public notebooks ranked by claimed score with the gap to our best
The notebook board: public notebooks ranked by the score they claim, the gap to ours, and the rounds that ran them

A local check is not the scorer. Two merges passed the quick screen and failed the full check, and one passed both and still errored on Kaggle. The record kept all three, so the same mistake is easy to spot when it comes back.

A better score can still mean a lower rank. The rank fell while the score improved every day. Watch the gap to the line you care about, not the rank alone.

What’s next

Submitting the 22.1580 portfolio, then real search on the larger n, where the best-known packings are least settled. The competition runs to July 2027. See all four runs in our November standings, or try the loop with Kaggle Overnight.

Try it on your next run

Every number in this post came from the ResearchForge record. Yours can too.