sutro-problems

RBF kernel ridge in a Fisher metric on a compressed training set: difficulty 2

Independent reviewer reproduction — October 19, 2026

The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs, with valid above-idle energy measurements (3 recorded GPU UUIDs). Median ranked time is 3.184645 ms/call and median energy is 207.682235 mJ/call; pooled MNIST error is 3.087576% (319,811/330,000 correct).

Reviewer run MNIST error Margin, labels Ranked ms Above-idle mJ
1 2.987273% +454 3.184645 189.029881
2 3.060000% +374 3.918747 207.682235
3 3.215455% +203 2.691631 281.843872

Source is unchanged at SHA-256 de16bb0926ae35895400758697ba68f73e7f5ec366f2b84495b74c0abba89804. 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.

This plain-PyTorch D2 port credits @sigkillme0’s accepted D3 Fisher/compression method, @spartanhaden’s earlier D2 kernel and @islamborghini’s kernel-PCG. The Fisher metric, log-radius feature, class-restricted compression and extrapolated PCG derive from #125; the triangular-solve preconditioner, side-stream factorization and D2 retuning are the submitted changes. Exact-source author records use working filename kk_d2_final.py and reconstruct separately. Development-call counts, probability estimates and optimization comparisons below were not independently reproduced by this review.

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-20261019/verify_records.py evidence/reviewer-20261019 submission.py --difficulty 2 --function kc_d2 --output /tmp/pr131-verification.json. This reconstructs recorded verdicts without executing the learner.

Original author report (as submitted)

Present-tense record and pending-review claims below describe the submission date; the dated reviewer section above gives the accepted review status.

Entrant: @spartanhaden.

This entry ports @sigkillme0’s merged d3 record (#125) to difficulty 2, in plain PyTorch, with a faster preconditioner solve and retuned constants.

Three sandboxed Modal A100-80GB runs of the unchanged scorer (mnist-a100/1.2.0, stock run_modal.py --difficulty 2 --runs 3) all passed.

Run GPU MNIST error Margin (labels) Worst draw Hold-out Ranked ms mJ above idle
1 A100-SXM4-80GB 3.165% +259 3.58% Fashion 86.76% 2.931 400
2 A100 80GB PCIe 3.113% +316 3.38% Fashion 86.67% 2.874 255
3 A100 80GB PCIe 3.023% +415 3.30% KMNIST 95.69% 2.699 282

Compared with the current record, our own rfm-pcg-d2 (4.6 ms, 603 mJ), this entry is 1.59× faster and uses 53% less energy per call.

submission.py is 16,400 bytes with SHA-256 de16bb0926ae35895400758697ba68f73e7f5ec366f2b84495b74c0abba89804, identical to source_sha256 in all three records (which name it by its working name, kk_d2_final.py). It passes check_source. evidence/official/manifest.json lists the hash of every file.

Method

No backprop. The whole call is one CUDA graph.

Our changes to #125’s method:

Together these save 0.72 ms of GPU time per call against our first port, at +0.018 points of error. Not used from #125: its Triton kernels, the natural-gradient step and the HKNN readout.

Certified before the official set on 1,056 fresh calls: mean error 3.109%, worst draw 3.71%, P(3/3) ≈ 0.995.

Credits

Notes

Reproduce from mnist-a100/:

python run_modal.py submissions/fisher-pcg-d2-20261019/submission.py:kc_d2 --difficulty 2 --runs 3 --json /tmp/d2-verification