The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs, with valid above-idle energy measurements on three distinct GPUs. Median ranked time is 219.194370 ms/call and median energy is 42,617.168403 mJ/call; pooled MNIST error is 2.273939% (322,496/330,000 correct).
| Reviewer run | MNIST error | Margin, labels | Ranked ms | Above-idle mJ |
|---|---|---|---|---|
| 1 | 2.283636% | +18 | 219.244261 | 42,847.432270 |
| 2 | 2.290909% | +10 | 219.194370 | 42,617.168403 |
| 3 | 2.247273% | +58 | 218.955311 | 40,462.843908 |
Source is unchanged at SHA-256 0bb2b7b4852db2bda7726029d7f211d109473d1b7c672a8256d6ce86daa94bef. The scorer and runner are unchanged. Verification reconstructs every complete canonical score and energy result from the raw records; source review covers per-call fitting, label isolation and cached-state reset. No reviewer outcome was replaced or retried.
The ensemble uses five deterministic cyclic orders of the same supplied training draw. Its RFM, kernel-PCG and HKNN source credits are preserved, including @spartanhaden’s accepted kernel teacher and RFM/HKNN implementation. Submitted external-test diagnostics remain separate from competition qualification.
The leaderboard uses the independent reviewer medians above. Author results below are retained separately. The published instructions specify three official Modal runs and their median; this independent reproduction is review evidence, not a newly imposed second-batch submission rule.
Using Python 3.13 as in the canonical runner, recheck saved records from this submission directory with python evidence/reviewer-20261018/verify_records.py evidence/reviewer-20261018 submission.py --difficulty 4 --function kernel_ensemble --output /tmp/pr123-verification.json. This reconstructs recorded verdicts without executing the learner.
Present-tense record and pending-review claims below describe the submission date; the dated reviewer section above gives the accepted review status.
Non-neural RBF kernel ensemble for difficulty 4, with three analytic metric updates per model, Nyström-preconditioned conjugate-gradient solves and local affine-hull correction. All three initial energy-enabled sandboxed Modal A100-80GB runs of the unchanged scorer (mnist-a100/1.2.0) passed: 218.463697 ms/call median, 39,296.268 mJ/call median above idle, and 2.209697% pooled MNIST error across 33 MNIST calls.
Difficulty 4 requires mean MNIST error at most 2.30% within each run, individual-draw error at most 3.80%, and holdout accuracy at least 15%.
| Run | GPU | Ranked ms/call | Above-idle mJ/call | MNIST error | Worst draw | Holdout accuracy |
|---|---|---|---|---|---|---|
| 1 | A100-SXM4-80GB | 217.963516 | 39,296.268 | 2.216364% | 2.49% | Fashion 88.6300% |
| 2 | A100-80GB PCIe | 219.338938 | 39,088.839 | 2.242727% | 2.57% | KMNIST 96.9575% |
| 3 | A100-SXM4-80GB | 218.463697 | 43,198.217 | 2.170000% | 2.50% | Fashion 88.2675% |
Five classifiers use different cyclic orders of the labelled training rows. Each learns an RBF feature metric through three analytic average-gradient-outer-product updates, fitted on 3,000, 6,000 and 10,000 labelled rows. Each update uses 12 CG iterations and up to 2,048 gradient samples; 2,000-row kernel panels cap the latter two updates at 2,000 samples.
Each classifier fits kernel ridge on all 10,000 labelled rows with a 256-landmark Nyström preconditioner, followed by a refit that includes confident query rows carrying model-predicted labels. An HKNN affine-hull correction revisits the 512 lowest-margin query rows using 70 neighbours for each of their three candidate classes. The five class-probability vectors are averaged. Query truth is never available to the learner.
The complete fit and prediction sequence runs in a CUDA graph captured during the permitted foreign warm-up. Every replay copies the current inputs, initializes solver and metric state, and refits all five classifiers. Cached objects hold graph execution and scratch storage, not reusable fitted models. The implementation uses plain PyTorch matrix operations and analytic input derivatives, with no backpropagation.
0bb2b7b4852db2bda7726029d7f211d109473d1b7c672a8256d6ce86daa94bef, matching every raw record. Source/import validation passes at 11,300 bytes, with no review flags. Records identify three distinct A100-80GB GPU UUIDs.These are author-run qualification results submitted for independent maintainer source review and verification. No established-record star is requested before that review.
The frozen source was also evaluated with the same 10,000-label budget and 9×9 → whitened/rotated 60-dimensional release:
| Test data | Unique query images | Fits | Mean error | Mean GPU ms |
|---|---|---|---|---|
| Standard MNIST official test split | 10,000 | 3 | 2.113333% | 219.220 |
| QMNIST additional test digits | 50,000 | 5 | 2.382000% | 219.255 |
The standard test images are reused across three independently sampled training/release draws. QMNIST uses five disjoint 10,000-query cohorts from indices 10,000–59,999, excluding the reconstruction of the standard test set. All calls use fresh canonical sandbox workers with foreign-data warm-up. No parameters were tuned using these results.
These are external generalization checks, separate from competition qualification. QMNIST’s pooled error exceeds 2.30%; this submission does not claim the competition accuracy threshold on that external dataset. Raw results, source hashes and dataset provenance are included.
From mnist-a100/:
python run_modal.py submissions/rfm-ensemble-d4-20261017/submission.py:kernel_ensemble \
--difficulty 4 --runs 3 --json /tmp/rfm-ensemble-d4
Verify the packaged qualification records:
python submissions/rfm-ensemble-d4-20261017/verify_records.py \
submissions/rfm-ensemble-d4-20261017/submission.py \
submissions/rfm-ensemble-d4-20261017/evidence/official
Source, report, raw qualification evidence, held-out results.