sutro-problems

Fisher-metric kernel ridge with a standard-form Woodbury preconditioner: difficulty 2

Independent reviewer reproduction — October 22, 2026

The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs. Median ranked time is 2.924269 ms/call; median above-idle GPU board energy is 172.382541 mJ/call. Pooled MNIST error is 3.109697% (319,738/330,000 correct). Each complete run independently passes all accuracy, holdout and timing gates; pooled accuracy is descriptive.

Reviewer run MNIST error Margin, labels Ranked ms Above-idle mJ
1 3.067273% +366 2.906669 322.672610
2 3.150909% +274 2.924269 171.027417
3 3.110909% +318 2.937826 172.382541

The submitted source is unchanged, SHA-256 311c1d980e162691e3872e7bafe18a357548d3c3442237e788c06570fcf6dde6; scorer and stock runner are unchanged. All 45 timed calls completed. Every initial outcome is retained; no maintainer-launched replacement or retry was made. Raw records and executable verifier, reconstruction, decision and source review are retained. Optional energy is assessed separately from score qualification and excludes host/system energy and initial compilation.

This is @spartanhaden’s linear-algebra optimization of the accepted Fisher-PCG D2. The report/source retain @sigkillme0’s Fisher/compression/PCG and @islamborghini’s kernel-PCG credits. Woodbury identities were checked independently; FP16 products can change numerical predictions. Author raw records reconstruct separately under working filename kd2_k.py. In-process call counts (1,056 in the PR body versus 2,112 in its README), pass-probability estimates, paired timing claims and earlier unsubmitted groups have no supplied raw development records and were not independently verified.

The leaderboard uses these independent reviewer medians; the original author report below remains separate. Published instructions specify three official runs and their median; this reviewer reproduction supplies evidence, without adding a permanent source exclusion or a new mandatory second-batch submission rule. No D6 qualification is claimed here.

Using Python 3.13 from this submission directory: python evidence/reviewer-20261022/verify_records.py evidence/reviewer-20261022 submission.py --difficulty 2 --function kc_d2 --output /tmp/pr136-verification.json. This audits saved records without executing the learner.

Original author report (as submitted)

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

Entrant: @spartanhaden.

This entry speeds up our merged fisher-pcg-d2 (#131). The learner and all its constants are unchanged; only the linear algebra and the stream layout around it are restructured.

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 80GB PCIe 3.083% +349 3.32% Fashion 86.83% 2.602 234
2 A100-SXM4-80GB 3.063% +371 3.51% KMNIST 95.59% 2.696 340
3 A100-SXM4-80GB 3.094% +337 3.32% Fashion 86.78% 2.986 147

Compared with the current record, our own fisher-pcg-d2 (reviewer median 3.18 ms, 208 mJ), this entry is 1.18× faster. Per-run energy varied from 147 to 340 mJ across hosts, and the median is above the record’s, so we make no energy claim.

submission.py is 17,238 bytes with SHA-256 311c1d980e162691e3872e7bafe18a357548d3c3442237e788c06570fcf6dde6, identical to source_sha256 in all three records (which name it by its working name, kd2_k.py). It passes check_source. evidence/official/manifest.json lists the hash of every file.

Method

The learner is the one in fisher-pcg-d2: an LDA + QDA Fisher metric with a log-radius coordinate, class-restricted k-means compression to 4,096 hard rows plus 240 weighted centroids, and RBF kernel ridge by 5 preconditioned CG steps plus an extrapolated direction. No backprop. The whole call is one CUDA graph.

What changed. Everything below is exact up to rounding, except the last item.

In-process on one A100-SXM4, paired over 24 draws, the call takes 2.31 ms against 2.66 ms for the merged file (−13%).

Certified on 2,112 fresh in-process A100 calls in two batches: mean error 3.117%, worst 11-draw split 3.39%, P(3/3) ≈ 0.99. Against the merged file the error rises by 0.002–0.004 points.

Credits

Notes

Reproduce from mnist-a100/:

python run_modal.py submissions/fisher-woodbury-d2-20261020/submission.py:kc_d2 --difficulty 2 --runs 3 --json /tmp/d2-verification

Co-developed with Claude.