The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs. Median ranked time is 78.333838 ms/call; median above-idle energy is 12,845.449793 mJ/call. Pooled MNIST error is 2.215455% (322,689/330,000 correct). Each run independently passes every accuracy, holdout and timing gate; pooled accuracy is descriptive.
| Reviewer run | MNIST error | Margin, labels | Ranked ms | Above-idle mJ |
|---|---|---|---|---|
| 1 | 2.197273% | +113 | 78.772494 | 13176.733304 |
| 2 | 2.192727% | +118 | 78.333838 | 12845.449793 |
| 3 | 2.256364% | +48 | 76.959902 | 11434.012357 |
The submitted source is unchanged, SHA-256 ec90dfb48793017577eaeac212b7357ff19bfef950dfb884025921b53198e608; scorer and stock runner are unchanged. All 45 timed calls completed. No outcome was replaced or retried. Raw records and executable verifier, reconstruction and source review are retained. Optional energy is assessed separately from score qualification.
This training-budget optimization retains @spartanhaden’s hybrid and its credits to @islamborghini, @likelytobelaura, @sjbaebae, RFM/HKNN and modded-nanogpt. Every function/class AST matches accepted #130. The sole executable changes are 400 to 250 student updates and target refresh chunks (13,27) to (8,17). Development summaries reconstruct separately; stochastic retrainings of the same draw are not independent draws.
The leaderboard uses these independent medians; original author results below remain separate. Published instructions require three official runs and their median; this reviewer reproduction supplies evidence and does not introduce a permanent source ban or an additional published submission rule.
Using Python 3.13, from this submission directory: python evidence/reviewer-20261019-pass3/verify_records.py evidence/reviewer-20261019-pass3 submission.py --difficulty 4 --function hybrid_d4 --output /tmp/pr132-verification.json. This audits saved records without executing the learner.
Pending-review and record claims below describe the author submission; the dated reviewer section gives accepted status.
Reducing the published tangent hybrid from 400 to 250 Adam updates, with cross-teaching moved from chunks (13, 27) to (8, 17), gives 77.010242 ms/call and 11,464.572860 mJ/call medians across three fresh canonical sandboxed Modal evaluations. All three runs pass difficulty 4 (MNIST error ≤2.30%); pooled MNIST error is 2.148182% (322,911/330,000 correct).
These are author execution results, not independent reviewer execution. Independent review is pending; this PR does not change the leaderboard. No failed qualification run was replaced or retried. Three distinct recorded GPU UUIDs are included.
| Run | MNIST error | Margin, labels | Ranked ms | Above-idle mJ |
|---|---|---|---|---|
| 1 | 2.160909% | 153 | 76.822557 | 11464.572860 |
| 2 | 2.051818% | 273 | 77.038087 | 11928.423796 |
| 3 | 2.231818% | 75 | 77.010242 | 11388.266687 |
Compared with the published hybrid’s independent reviewer medians (97.047379 ms, 14,754.131810 mJ), this author group is 20.65% faster and uses 22.30% less above-idle energy. These are separate draws and hosts, not a paired accuracy comparison. The parent’s pooled reviewer error was 2.083333%; reducing training trades some accuracy margin for speed.
The algorithm is @spartanhaden’s hybrid, unchanged apart from student update count, proportional cross-teaching schedule, and corresponding method documentation. The initial source SHA-256 is 102363cfe2825eee42b5ee641731655679825e7087c3d83e19c3340d93859306. This submission’s exact tested source is 20,458 bytes, SHA-256 ec90dfb48793017577eaeac212b7357ff19bfef950dfb884025921b53198e608; all three raw records identify those same bytes under their execution filename hybrid-d4-steps250-20261019.py. Packaging renames the file to submission.py without changing its contents.
Retained components: two RFM teachers with three AGOP metric updates, FlashAttention kernel products, Nyström-preconditioned CG, HKNN; three 60–256–256–10 RMSNorm/ReLU students; 1,024 labelled plus 1,024 pseudo-labelled query rows per update; local tangent jitter and isotropic noise; Adam and the original fractional LR anneal; teacher/student probability mixture and final HKNN. Every call refits the teachers and resets/retrains students. Only graphs, scratch tensors and compiled code persist. Query labels are never used.
The parent credits @islamborghini’s kernel-PCG/RFM ensemble, Radhakrishnan et al. RFM, Vincent and Bengio HKNN, @likelytobelaura’s tangent cross-teaching/ReLU ensembles, @sjbaebae’s graph-readout work and modded-nanogpt’s zero head initialization. All original source credits are retained. This is a training-budget optimization of their learner, not a new learner or kernel implementation.
Before qualification, eleven fresh paired development input draws, each with three stochastic retrainings, gave 2.214242% mean error and 75.396596 ms warm full-call wall time at 250 updates; the unchanged 400-update parent gave 2.115758% and 92.884268 ms. Those development timings exclude sandbox overhead and are separate from the canonical results above. The 200-update alternative had weaker mean accuracy margin (2.274848%); it was not substituted into qualification. Repeated stochastic trainings on one input draw are not independent input draws.
From mnist-a100/, run the exact method with the unchanged runner and scorer:
python run_modal.py submissions/hybrid-tangent-250-d4-20261019/submission.py:hybrid_d4 --difficulty 4 --runs 3 --json /tmp/hybrid250-d4-verification
The saved raw records include all initial outcomes, complete call data, energy windows, GPU identities and source hashes. Verification reconstructs every verdict, ranked time, summary and energy value through the repository’s canonical judge and energy_summary. The verifier is adapted from the parent’s published reviewer verifier; it audits saved evidence and does not constitute another fresh execution group.
From this submission directory:
python verify_records.py evidence/official submission.py --difficulty 4 --function hybrid_d4 --record-filename hybrid-d4-steps250-20261019.py --output evidence/official/verification.json
The scorer SHA-256 is 1bc6d8d96a776565bc70f3ccdb91be7de57187d4495b2b0437fac98f5e120064 and runner SHA-256 is 1180b843cf6c32414774da90c835a8b7298d2db8719989237a793ca1c62a10ab, byte-identical to the repository versions used for this audit. Compilation/graph capture happens during canonical warm-up; ranked measurements cover complete subsequent calls, including teacher fitting, training and readout. The unchanged warm-up and timing checks passed in every run. Source review can be performed by comparing this file against the linked parent; no custom native-kernel experiments are included.