sutro-problems

D5: cross-taught Muon MLP ensemble with local tangent jitter

Entrant: @likelytobelaura. The joint train+query graph-propagation readout is @sjbaebae’s, from the D3 record (see Credits).

Three canonical Modal A100-80GB sandbox evaluations, the first and only set run for this source, all passed difficulty 5 (mean MNIST error at most 1.90%). Median ranked time is 3,359.445 ms per call; median energy above idle is 677,766.5 mJ per call. The 33 MNIST calls labelled 324,184 / 330,000 rows correctly (1.762% aggregate error).

Run GPU MNIST error Margin, labels Hold-out Ranked ms mJ above idle Record
1 A100 80GB PCIe 1.790% +121 KMNIST 97.63% 3,369.836 633,927 json
2 A100-SXM4-80GB 1.748% +167 Fashion 88.93% 3,359.445 677,766 json
3 A100-SXM4-80GB 1.749% +166 KMNIST 97.52% 3,352.078 699,197 json

The three runs used three distinct GPU UUIDs. The worst single MNIST draw was 2.06% (the cap is band + 1.5 = 3.40%), and every call was within 0.7% of its dataset’s median time. verification.json recomputes every verdict, the ranked times and the per-call gates from the raw per-call data in the records.

Source: submission.py, 12,135 bytes, SHA-256 a3b18bac64dddc3f7aa8c614e2279932aa74040720cb861e8dfa40eaa67923c3 (identical to the source_sha256 in all three records). mnist.check_source passes and mnist.review_flags reports nothing. The file is source only: about 35 scalar hyperparameters, no arrays, no fitted data.

Method

Rules review

Checked against the rules in mnist.py’s docstring:

Development (A100, dev/quick.py, not official scores)

Step Mean error, seeds 401-408 ms per call
D4 cross-taught ensemble (1,700 updates) 2.149% 2,170
+ signed-power embeddings, temperature 20, final-readout hub correction 2.070% 2,163
+ local tangent jitter, strength 1.0 1.847% 2,343
+ 2,500 updates 1.801% 3,383
+ no input noise 1.785% 3,383
This file, seeds 401-416 1.793% 3,497

With tangent jitter on, these were neutral or worse: middle-layer embeddings, 16 members, width 384, more than 2,500 updates, 5 or 20 neighbours, 2 or 8 directions, no mixup, and tangent jitter of query rows (label-free or along pseudo-class neighbours).

Credits

Reproduce

From mnist-a100:

python run_modal.py submissions/tangent-crossteach-d5-20261012/submission.py:classify --difficulty 5 --runs 3 --json out/d5-tangent

Independent maintainer verification — 2026-10-12

All three initial fresh energy-enabled canonical Modal A100-80GB runs passed. Source review found no compliance blocker: caller inputs are copied into owned buffers, tangents are fitted anew, and weights, EMA, optimizer state, pseudo-labels and query pools are reset on every call. All 45 timed calls completed and three distinct GPU UUIDs were recorded. No failed outcome was replaced, and no learner retry or repair was made.

The reviewer set has median ranked time 3335.394886 ms, median above-idle GPU board energy 638234.154568 mJ, and descriptive pooled MNIST error 1.824848%. Each run passes the 1.90% gate individually. The top-level leaderboard uses this reviewer set; author and research records remain separate.

Reviewer run Correct MNIST error Ranked ms Above-idle mJ
1 107,940/110,000 1.872727% 3335.394886 662154.992877
2 108,063/110,000 1.760909% 3358.272278 629924.834404
3 107,975/110,000 1.840909% 3322.150457 638234.154568

Raw records and frozen manifest, canonical reconstruction, and source review are retained. The learner SHA remains a3b18bac64dddc3f7aa8c614e2279932aa74040720cb861e8dfa40eaa67923c3.

Verify the reviewer records locally from mnist-a100/:

python submissions/tangent-crossteach-d5-20261012/evidence/reviewer-20261012/verify_records.py \
  submissions/tangent-crossteach-d5-20261012/evidence/reviewer-20261012 \
  submissions/tangent-crossteach-d5-20261012/submission.py \
  --output /tmp/d5-reviewer-verification.json

Energy is above-idle GPU board energy under the canonical repeated-call protocol, excluding host/system energy and first-call compilation. Comparisons with other submissions use separate draws and devices, not a paired accuracy experiment.