sutro-problems

Two RFM kernel teachers distilled into tangent-jittered MLP students: difficulty 4

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 97.047379 ms/call and median energy is 14,754.131810 mJ/call; pooled MNIST error is 2.083333% (323,125/330,000 correct).

Reviewer run MNIST error Margin, labels Ranked ms Above-idle mJ
1 2.127273% +190 97.047379 14,754.131810
2 2.107273% +212 97.567776 15,407.721624
3 2.015455% +313 94.964916 13,283.403602

Source is unchanged at SHA-256 102363cfe2825eee42b5ee641731655679825e7087c3d83e19c3340d93859306. 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 teacher credits @spartanhaden’s RFM/HKNN learner, @islamborghini’s kernel-PCG and RFM ensemble, including cyclic row orders, the third metric update and softmax averaging. The student design credits @likelytobelaura’s tangent cross-teaching and ReLU/Muon ensemble; those upstream reports credit @sjbaebae for graph propagation. The submitted report also credits the kernel-taught feature/target design and modded-nanogpt’s zero head initialization, and distinguishes held #126’s related design. Exact-source author records use working filename sd4_z400.py. Their results reconstruct separately. Development-call counts and probability estimates 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 4 --function hybrid_d4 --output /tmp/pr130-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.

Three sandboxed Modal A100-80GB runs of the unchanged scorer (mnist-a100/1.2.0, stock run_modal.py --difficulty 4 --runs 3) all passed.

Run GPU MNIST error Margin (labels) Worst draw Hold-out Ranked ms mJ above idle
1 A100-SXM4-80GB 2.106% +213 2.28% Fashion 88.67% 101.302 14,893
2 A100-SXM4-80GB 2.110% +209 2.33% KMNIST 96.80% 100.563 14,573
3 A100-SXM4-80GB 2.095% +225 2.26% KMNIST 96.63% 97.944 14,943

Compared with the current record (219.2 ms, 42,617 mJ; @islamborghini’s RFM ensemble), this entry is 2.18× faster and uses 65% less energy per call.

submission.py is 20,460 bytes with SHA-256 102363cfe2825eee42b5ee641731655679825e7087c3d83e19c3340d93859306, identical to source_sha256 in all three records (which name it by its working name, sd4_z400.py). It passes check_source. evidence/official/manifest.json lists the hash of every file.

Method

Plain PyTorch. Each call trains from scratch, and the whole call is one CUDA graph captured in the warm-up.

Certified before the official set on 4,224 fresh in-process A100 calls: mean error 2.116%, P(3/3) ≈ 0.99.

Credits

Notes

Reproduce from mnist-a100/:

python run_modal.py submissions/hybrid-tangent-d4-20261019/submission.py:hybrid_d4 --difficulty 4 --runs 3 --json /tmp/d4-verification