sutro-problems

Fisher-metric kernel ridge with an HKNN readout, in plain PyTorch: difficulty 3

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 3.510784 ms/call; median above-idle GPU board energy is 358.758434 mJ/call. Pooled MNIST error is 2.523030% (321,674/330,000 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 2.557273% +157 3.510784 444.613117
2 2.509091% +210 3.380503 358.758434
3 2.502727% +217 5.392590 351.673144

The submitted source is unchanged, SHA-256 eae98d2337442db3c2ed5984e455cf2fc945bfa6846aeaa93d0decb51799d232; 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.

The report/source credit @sigkillme0’s Fisher/compression/PCG method, @spartanhaden’s RFM/HKNN and @islamborghini’s kernel-PCG, including Vincent and Bengio’s hull method. This is a whole-learner comparison with the previous native-hull record, rather than an isolated kernel experiment. Current labels/metrics, preconditioner and compacted low-margin readout are refitted each call. Author raw records reconstruct separately under working filename kd3_k.py; in-process counts, probability estimates, paired screens and unsubmitted source histories lack supplied development records and remain unverified.

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 3 --function kc_d3 --output /tmp/pr137-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.

Our plain-PyTorch d3 learner: @sigkillme0’s #125 method, re-implemented, with the HKNN readout of our rfm-hknn-d3 and our d2 entry’s linear algebra. No hand-written kernels.

Three sandboxed Modal A100-80GB runs of the unchanged scorer (mnist-a100/1.2.0, stock 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 2.541% +175 2.78% Fashion 86.71% 4.183 614
2 A100-SXM4-80GB 2.450% +275 2.71% KMNIST 96.05% 3.537 468
3 A100-SXM4-80GB 2.530% +187 2.85% Fashion 87.00% 3.716 389

Compared with the current record, @sjbaebae’s native-hull-cg-d3 (#134; reviewer median 4.558 ms, 639 mJ), this entry is 1.23× faster and uses 27% less energy per call.

submission.py is 20,453 bytes, 27 under the size limit, with SHA-256 eae98d2337442db3c2ed5984e455cf2fc945bfa6846aeaa93d0decb51799d232, identical to source_sha256 in all three records (which name it by its working name, kd3_k.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.

Changes from our 10-18 port kc_d3_v3m: the d2 changes (side-stream class moments, Woodbury class covariances and preconditioner with fp16 products, batched small Choleskys, L⁻ᵀL⁻¹ precision, one operand padded from 99 to 104 columns); a max-based top-2 margin; and HKNN rows taken as the first 512 with margin below 0.2 by a cumsum compaction, not the 512 lowest (+0.002 points).

In-process on one A100-SXM4 (24 paired draws), the call takes 3.34 ms, against 4.46 ms for #134 (−25%) and 4.01 ms for our 10-18 port.

Certified on 1,056 fresh in-process A100 calls: mean error 2.432%, worst 11-draw split 2.64%, P(3/3) ≈ 0.99. The official runs averaged 2.507%, within noise for 33 draws.

Credits

Notes

Reproduce from mnist-a100/:

python run_modal.py submissions/fisher-hknn-d3-20261020/submission.py:kc_d3 --difficulty 3 --runs 3 --json /tmp/d3-verification

Co-developed with Claude.