Two 60-256-256-10 MLPs train for 30 updates in a CUDA graph. The softmax heads train the embeddings; three-neighbor voting over labelled training embeddings makes the final prediction. Native CUDA and cuBLAS implement the computation. The standalone source embeds its CUDA code.
Three of three fresh official Modal A100-80GB runs passed for the unchanged source. Median ranked time is 6.705614625 ms per call; median energy above idle is 529.576891 mJ per call. The 33 MNIST calls together predicted 314,013 / 330,000 labels correctly, for 4.844545% aggregate error, below the difficulty-1 bound of 5.40%.
| Run | GPU / power limit | MNIST error % | Ranked ms | Energy mJ above idle | Raw evidence |
|---|---|---|---|---|---|
| 1 | A100 80GB PCIe, 300 W | 4.898182 | 6.705615 | 529.576891 | JSON |
| 2 | A100-SXM4-80GB, 400 W | 4.877273 | 7.178109 | 501.677797 | JSON |
| 3 | A100-SXM4-80GB, 400 W | 4.758182 | 6.131708 | 760.733904 | JSON |
The canonical runner sampled three distinct boards: one PCIe and two SXM4. The reported median includes all three initially requested runs, with no retries or result selection. Each run used 11 MNIST and four KMNIST holdout calls, with fresh sandboxed workers and a foreign warmup for each timed call. All canonical accuracy, timing, sandbox and energy gates passed. Official software was PyTorch 2.12.0+cu130 with scorer mnist-a100/1.2.0.
The ranked time includes the scorer’s independent parent-process clock floor, after its empty-method overhead adjustment; CUDA-event time alone does not determine the score. For example, run 1 averaged 3.426769 ms of MNIST CUDA-event time and 3.475200 ms on KMNIST, while its canonical ranked time was 6.705615 ms. The historical SSH-hosted measurements below used another environment and do not replace this official result.
The exact source is SHA-256 d165b1cb8c395e3b2f3aa2f228c101818eaf389893b271b5d6d15358aaba8987, 20,422 bytes. Source identity, every canonical judge result, ranked time and full energy record were independently reconstructed from the saved records in verification.json. Execution provenance is in manifest.json; exact copied evidence hashes are in evidence-sha256.json. The learner, author local records and author validation manifest are unchanged.
All three evaluations passed for the exact same source: 3.563617 ms median ranked time, 4.864545% mean MNIST error across evaluations. Each evaluation used 11 fresh MNIST calls and four holdout calls in fresh sandboxed processes (330,000 MNIST predictions total). Sandbox was explicitly required; all records report sandbox enabled and no problems.
| Evaluation | MNIST error % | Ranked ms | Raw evidence |
|---|---|---|---|
| 1 | 4.939091 | 3.451706 | JSON |
| 2 | 4.839091 | 3.665033 | JSON |
| 3 | 4.815455 | 3.563617 | JSON |
Source SHA-256: d165b1cb8c395e3b2f3aa2f228c101818eaf389893b271b5d6d15358aaba8987; size 20422 bytes (limit 20,480). The source directory now includes the process UID to avoid a permission collision with a preexisting development compilation directory. Learning and native kernels are unchanged by that loader fix.
Device: NVIDIA A100-SXM4-80GB. PyTorch: 2.12.1+cu129. Evaluator: mnist-a100/1.2.0. The upstream established D1 time at base a2ed895 is 61.7 ms; the local median is 17.31 times faster.
These three SSH-hosted sandbox evaluations preceded the official qualification. Energy was skipped in these local runs. Their records and validation manifest remain historical evidence; the leaderboard now uses the official three-run medians above.
From mnist-a100, with Modal configured, reproduce the canonical three-run timing and energy protocol:
python run_modal.py submissions/native-mlp-d1-20261007/mlp_neighbor3_register_d1.py:classify \
--difficulty 1 --runs 3 --json /tmp/native-mlp-d1-official
To reproduce a historical local evaluation in a fresh process on an A100 with CUDA development tools and a working sandbox installation:
MNIST_SANDBOX=required python mnist.py submissions/native-mlp-d1-20261007/mlp_neighbor3_register_d1.py:classify --difficulty 1 --no-energy --json evaluation.json
Repeat the local command three times with distinct output paths. Both protocols randomly release fresh draws each time; the official command also measures the canonical energy column.