sutro-problems

Learned-metric kernel ridge with HKNN on low-margin rows: difficulty 3

Entrant: @spartanhaden.

Three sandboxed Modal A100-80GB runs of the unchanged scorer (mnist-a100/1.2.0, run_modal.py --difficulty 3 --runs 3) all passed.

Run GPU MNIST error Margin (labels) Worst draw Hold-out Ranked ms mJ above idle
1 A100-SXM4-80GB 2.429% +298 2.64% KMNIST 96.54% 18.193 2,301
2 A100-SXM4-80GB 2.399% +331 2.65% Fashion 88.16% 19.228 2,218
3 A100-SXM4-80GB 2.485% +236 2.63% KMNIST 96.62% 18.391 2,181

Compared with the current record (51.8 ms, 8,839 mJ; @sjbaebae, spherical metric kernel (#117); #116 reports 51.6 ms), this entry is 2.8× faster.

submission.py is 14,707 bytes with SHA-256 da456d0ed99d8fb59b26e76b1658d6cd8d69b31710122936f2ca5f81f0ced966, identical to source_sha256 in all three records. It passes check_source. evidence/official/manifest.json lists the hash of every file.

Method

Credits

Builds on @islamborghini’s kernel-pcg entry and Recursive Feature Machines (Radhakrishnan et al. 2022). The local-hull step is HKNN (Vincent & Bengio, NeurIPS 2002).

Notes

Independent maintainer verification — 2026-10-12

All three initial fresh energy-enabled sandboxed Modal A100-80GB runs passed individually. Source review found no qualification blocker. All 45 timed calls and complete energy windows were reconstructed with the unchanged canonical scorer; three distinct GPU UUIDs were recorded. No failed run was replaced or learner repair made.

The reviewer medians are 18.143665 ms and 2361.815112 mJ above idle per call. The leaderboard uses this independent set; the author source and original evidence above are retained.

Reviewer run Correct MNIST error Ranked ms Above-idle mJ
1 107,400/110,000 2.363636% 18.411804 2100.016333
2 107,256/110,000 2.494545% 18.112284 2361.815112
3 107,212/110,000 2.534545% 18.143665 2370.329065

Raw records and manifest and canonical reconstruction are retained. Learner SHA-256: da456d0ed99d8fb59b26e76b1658d6cd8d69b31710122936f2ca5f81f0ced966. Energy excludes host energy and first-call compilation.

Reproduce from mnist-a100/:

python run_modal.py submissions/rfm-hknn-d3-20261012/submission.py:rfm_d3 --difficulty 3 --runs 3 --json /tmp/pr121-d3-verification