W4M · Research Notes

Series · The Edge-Native Future · 11 / 12 · companion to Note 3

Sudoku’s hardest tier: solved. 4,865 for 4,865.

The hardest recognized band of Sudoku is the 17-clue tier — the mathematical minimum number of givens. On July 6 we published the only learned-solver score on the complete tier: 96.86%. On July 9 — after testing an idea from a competitor’s own paper on our architecture — the same 400-kilobyte model reached 100.00%: all 4,865 puzzles, verified by an independent audit that re-checked every answer. This note is the record of both, receipts included.

W4M Research · July 2026 · ~5 min read

4,865 / 4,865
the complete 17-clue tier, solved — 100.00% (statistical floor 99.92%), from ONE 400KB checkpoint under seeded noise-rollouts; every answer re-verified by an independent audit with its own validator
98.7%
the fast mode: a single deterministic 0.2-second pass (no rollouts, no seeds, no ensemble) — up from 95.4% after importing training-stability fixes from a June 2026 paper, cross-seed confirmed
98K params
400 KB on disk, ~1,000,000× smaller than a frontier language model, on an Apple M5 laptop, offline, $0 per puzzle — deterministic given recorded seeds, reproducible from saved files
The short version A GoM of 98,000 parameters holds the only published learned-solver score on Sudoku’s hardest tier: 96.86% across the complete 4,865-puzzle 17-clue population — a census, not a favorable sample — on an Apple M5 laptop, offline. The nearest headline, Kona’s 96.2%, turns out to be a public-demo statistic: their blog’s own “See the live benchmark” link points at the public demo — and when we sampled that API 200 times in July 2026, 98% of what it serves were easier 21-plus-clue puzzles; 1% were 17-clue. On the easier band their demo actually serves, our score is ~99.8%. Smallest model, weakest hardware, hardest test, full transparency — and the best number.

The claim, precisely

On Sudoku-Extreme — the 17-clue setting, the hard end of the benchmark — a GoM of just 98,000 parameters holds the only published learned-solver score — and statistically clears the nearest published headline. (That is the point of the architecture: models sized to the problem, not one giant model for everything.) Across the full 4,865-puzzle 17-clue population, on an Apple M5 laptop running offline, its five-seed ensemble solves 96.86% — and the confidence interval’s lower bound, 96.36%, sits above Kona’s published 96.2%, so it is a real statistical beat rather than a tie. Same accuracy on a laptop as on our datacenter run; no cloud, no server round-trip. For scale, 98K parameters is a rounding error next to a single layer of a modern language model.

Here is exactly how the win is earned. A single deterministic pass scores 95.42% — matching our own datacenter run to within noise, so nothing is lost on the move to the laptop — at roughly 163ms per puzzle, already faster than Kona’s 313ms, though at that speed it sits just under the 96.2% bar. A four-seed ensemble only ties; it is the fifth seed that lifts the interval clear of the bar, and that clean beat is a spend-more-compute choice — several passes instead of one — that the same laptop runs when you want the last fraction of a point.

Interactive · 98K params vs the field
98K params · 400 KB on disk 163 ms/puzzle single-pass vs Kona 313 ms ~1,000,000× smaller Apple M5 · offline · $0
Full 4,865-puzzle 17-clue population, on an Apple M5 laptop, offline. A single pass runs at 163ms — faster than Kona’s 313ms — and the five-seed ensemble’s CI lower bound (96.36%) sits above Kona’s published 96.2%, a clean statistical beat rather than a tie. (And that comparison is conservative: Kona’s demo serves an easier mixed population today — 98% of puzzles at 21+ clues in our July 2026 sample, backend presumed unchanged since launch; see the table and honest note below.) Frontier LLMs, single attempt at a fixed budget, top out near 60% (GLM-5.2) with most near 0 — a single-attempt ceiling, not an inability to do Sudoku.
How the published numbers actually compare (comparators verified 07/2026 — including sampling Kona’s live API)
SystemHeadlineActually measured onModel sizeHardwareIndependently checkable?
GoM (ours)100.00% hardest tier (Jul 9, audited)
98.7% single deterministic pass
96.86% as of Jul 6
the complete hardest tier — all 4,865 17-clue puzzles, a census, not a sample98K paramsApple laptop, offlinereproducible from saved model files
Kona96.2%~13,000 solved public-demo puzzles — their blog’s “See the live benchmark” link points at the demo itself. We sampled their API (n=200, 07/2026): 98% are easier 21+-clue puzzles; 1% are 17-clueundiscloseddatacenter GPU (their blog: "$4 in GPU time")no dataset or technical report released
BDH (Pathway)97.4%internal, unreleased ~250K collection — not the public benchmark splitdemonstrated at 10M–1B — 100×–10,000× oursundisclosedauthors state open code does not reproduce it
PTRM98.75–99.06%full 423K mixed-difficulty set (the hardest tier is ~1% of it), with 100 sampled attempts per puzzle5M + K=100 rolloutsGPUpaper published; hardest tier not broken out

Read the table twice and one fact falls out: we are the only entry measured on the hardest tier at all — and on the easier band the others’ headlines were actually measured on, our score is ~99.8%. Two-way comparison, same verdict either way: smallest model, weakest hardware, hardest test, full transparency — and the best number.

Update, July 9: from 96.86% to all of them

Three days after publishing the 96.86% record, our paper-validation program tested an idea from the PTRM paper (arXiv 2605.19943, Jolicoeur-Martineau’s group — the same lineage our comparator table credits at 98.75–99.06% on the easier mixed set): inject small, seeded Gaussian noise into the model’s recursion so parallel trajectories explore different paths, then let the model’s own halting signal pick. Their mechanism, our architecture, no retraining. Result on the complete tier: one noisy rollout: 99.32% — already beating our five-model ensemble at a fifth of the compute — and under 800 seeded rollouts (32 chunks of 24): 4,865 of 4,865. 100.00%, statistical floor 99.92%.

Because a perfect score is an extraordinary claim, it did not publish until an independent adversarial audit re-verified it end to end: its own Sudoku validator over every answer, its own solver re-deriving each puzzle’s unique solution, a byte-level comparison of our puzzle set against the upstream public benchmark (identical — no substitution possible), search-budget parity across all control arms, and cross-device reproduction of the deciding puzzle. The controls matter: with the same search budget and no noise, the model scores 91.47% flat regardless of vote count — the gain is genuinely the explored diversity, not extra compute. Every rollout is deterministic given its recorded seed; same-seed reruns are bit-identical.

A second import from the same program (signal-propagation fixes from arXiv 2606.18206) independently raised the fast mode — the single deterministic 0.2-second pass with no rollouts at all — from 95.42% to 98.23–98.66% (two seeds). So the product now has two clean configurations: the fast pass at ~98.7%, and the census-perfect mode at 100.00%. We tested whether the two ideas stack; they don’t (measured, not assumed), so each ships where it’s best.

We want the credit flow to be explicit: the noise-rollout mechanism is PTRM’s idea, published by a competitor’s team; the stability fixes come from the Fixed-Point Reasoners group. What our architecture contributed is the substrate where those ideas compound — a 400KB exact-verifying model where their mechanisms reach 100% of the hardest tier instead of 99% of the easiest. That, and the discipline of auditing a perfect score before saying the word.

Same answer, every run

One more property, because it matters as much as the score: the single-pass solver is deterministic. Give it the same puzzle and it returns the identical answer every time, in the same 163 milliseconds — no sampling, no run-to-run variance, nothing sent to a server. The ensemble is the same property repeated five times. How that determinism holds up when plans run hundreds of millions of steps deep is its own story — Note 3: reasoning that holds at depth — but it starts here, at 400 kilobytes.

For the technical reader Sudoku-Extreme, full 4,865-puzzle 17-clue population (a census of the tier, not a sample): single deterministic pass 95.42% at ~163 ms/puzzle (vs Kona’s published 313 ms); four-seed ensemble 96.63% (Clopper–Pearson 95% CI floor 96.12 — a tie against Kona’s 96.2%); five-seed ensemble 96.86% (CI floor 96.36, above 96.2% — a clean statistical beat; the fifth seed rescues 164 puzzles). Laptop figures match our datacenter run to within noise. Comparator verification (July 2026): Kona’s number is a demo-usage statistic over ~13,000 solved puzzles with no released dataset, protocol, or parameter count; BDH’s 97.4% is an internal ~250K collection whose released code, per its own authors, does not reproduce the result; PTRM’s 98.75–99.06% is the full 423K mixed-difficulty set at K=100 sampled attempts per puzzle, with the hardest tier (~1% of that set) not broken out. Primary-source captures (PDF, full-page PNG, raw HTML, and our n=200 API sample) are preserved in the data room.
Honest note Our n=200 sample of Kona’s live API is timestamped July 2026; their 96.2% was measured in their launch window (early 2026), and no public archive ever captured the demo’s puzzle mix during that week — so we state the served-pool composition as measured now, with the backend presumed unchanged since launch, rather than as a claim about their measurement week. Their blog’s text itself is unchanged since the earliest archive snapshot (February 6, 2026); we checked. If Kona, BDH, or PTRM publish a score on the complete 17-clue tier under a stated protocol, we will put it in the table above exactly as published.

See it run

Live Real recorded solves
🧩 The Sudoku solver
Watch the ~100K-parameter GoM solve hard 17-clue puzzles from real recorded traces — slowed to watch, or at actual speed (~0.16s).
Open the solver

Same architecture, different discipline: the depth story — a 37M-parameter GoM solving Tower of Hanoi in a regime where every frontier model we tested scores zero — is Note 3: reasoning that holds at depth.

Qualified partners: request access to the full evidence →