Research inclusion; D5 reproduction remains unresolved. The original author group below passed 3/3 (median 2,580.846 ms). The separate initial reviewer group passed 1/3: two runs stopped at later foreign-data warm-ups after seven timed calls each. Their partial accuracies do not qualify incomplete runs; the underlying stall cause remains undetermined. All outcomes are retained, with no qualifying reviewer-group median or leaderboard promotion. The source, original author evidence and credits below are preserved. See reviewer evidence.
This submission is a credited derivative of @likelytobelaura’s PR #118, pinned at 436b80bb40e4c346792ddad71670c7a5dbc8c5e9. It reduces training to 2,000 updates, refreshes cross-teaching pseudo-labels every 400 updates, and lets both groups sample all query rows rather than disjoint query halves. The original joint train/query graph-propagation readout is @sjbaebae’s. Muon and its Newton–Schulz coefficients are credited to Keller Jordan et al.
All three original official Modal A100 sandbox evaluations passed D5. Median ranked runtime is 2,580.846 ms, median energy above idle 536,878.610 mJ, and pooled MNIST error 1.848182% (323,901 / 330,000 correct). These are measured submission results, not a leaderboard update. No qualification reruns replaced failures.
| Run | MNIST error | Margin in labels | Ranked ms | Energy mJ |
|---|---|---|---|---|
| 1 | 1.827273% | +80 | 2580.846 | 536878.610 |
| 2 | 1.824545% | +83 | 2577.045 | 514395.945 |
| 3 | 1.892727% | +8 | 2724.036 | 562568.582 |
The tightest run has only eight labels of margin. Passing this fixed three-run evaluation does not establish that every future batch will pass. Raw records preserve scorer stdout, stderr, all 15 timed calls, energy windows, device UUIDs, sandbox status and source hash. verify.py checks their consistency against the unchanged source.
Eight batched MLPs, 60–256–256–256–10, use RMSNorm/GELU and dropout 0.1. Hidden matrices use Muon (three Newton–Schulz steps); gains, biases and output heads use AdamW. Training uses a cosine schedule, EMA, and mixup Beta(0.5,0.5). Each labelled row has four local tangent directions computed from its ten nearest same-class training rows; fresh Gaussian tangent jitter is applied online. True-label batches contain 2,048 rows. From update 400, batches also contain 1,024 pseudo-labelled query rows. The two halves of four models cross-teach at updates 400, 800, 1200 and 1600, selecting their top 90% confident query predictions. Both halves sample all queries. There is no separately trained teacher/student stage.
The final readout centers last-hidden EMA embeddings, applies signed power 0.25, normalizes and concatenates members, then uses a K5 labelled-neighbor prior and three joint train/query graph-propagation steps. Final similarities use hub correction 0.3. Training is BF16-oriented compiled PyTorch with CUDA graphs, not the earlier native BN teacher/student kernel implementation.
Each call recomputes tangents and initializes weights, optimizer state, EMA, query pools and pseudo-labels from scratch. Only the supplied train_x, train_y and unlabelled test_x enter the learner; there is no offline fitted data or query-label access.
comparison-2500/ contains the separately frozen 2,500-update, every-400-refresh variant and its three original official evaluations. All passed: errors 1.802727%, 1.800909%, 1.767273%; pooled error 1.790303%; median 3,385.827 ms and 674,307.053 mJ. This is a slower accuracy-margin comparison, not the submitted timing.
research/paired-development.json contains eleven paired fresh development draws. The 2,000-step one-refresh control scored 1.747273% at 2484.646 ms; periodic refresh scored 1.723636% at 2579.867 ms. At 2,500 steps, one refresh scored 1.699091% at 3249.113 ms and periodic refresh 1.668182% at 3368.803 ms. These development draws were easier than the official set; their absolute accuracy is not a qualification claim.
The earlier 2,000-step one-refresh candidate passed only 2/3 official evaluations: 1.880000%, 1.954545%, 1.799091%. All initial records and its source are retained in research/once2000/. It is not qualified and is not substituted for this submission.
From mnist-a100:
python run_modal.py submissions/tangent-refresh400-d5-20261012/submission.py:classify --difficulty 5 --runs 3 --json out/d5-refresh400
python submissions/tangent-refresh400-d5-20261012/verify.py
submission.py is 10,883 bytes, SHA-256 6f0e1a109ece2679f8ee1e0527cd98abc6a58e664d11fc6033740765ba4a990f, matching every primary official record. The 2,500-step comparison hash is 86d7caab67e4c9de30a698652fab6195599e5184113c359a83872f5135851018. Both executable ASTs were verified against the development variants before launching official evaluation.