The repaired source passed all three initially requested fresh canonical Modal A100-80GB runs, including full energy measurements. Median ranked time is 416.134586331 ms per call; median GPU energy above idle is 98,226.023073203 mJ per call. The 33 MNIST calls predicted 321,509 / 330,000 labels correctly, for 2.573030% aggregate error. Each individual run passed the difficulty-3 2.70% mean-error gate and every other canonical timing, holdout, sandbox and energy gate.
| Run | GPU | MNIST error % | Margin, correct labels | Ranked ms | Energy mJ above idle | Raw evidence |
|---|---|---|---|---|---|---|
| 1 | NVIDIA A100 80GB PCIe | 2.573636 | +139 | 384.850889 | 83843.534463 | JSON |
| 2 | NVIDIA A100-SXM4-80GB | 2.479091 | +243 | 416.134586 | 98226.023073 | JSON |
| 3 | NVIDIA A100-SXM4-80GB | 2.666364 | +37 | 420.447074 | 100261.515947 | JSON |
All three runs used distinct GPU UUIDs and completed 11 MNIST and four holdout calls. Ranked time includes the unchanged scorer’s parent-clock floor; GPU energy is idle-adjusted board energy, not whole-system energy. These observed passes do not estimate future qualification probability.
The submitted learner remains byte-identical at SHA-256 c6082659a49a6bd942077e359db42f28016626bb49303d68032467710d765a7e, 17,973 bytes. The source-only archive and all 15 decoded files were reviewed for label isolation and reset of weights, momentum and SWA on every graph replay; no fitted-data or retained-trained-state blocker was found. Human source review records the resolved encoded-source advisory flags.
The unchanged scorer and runner identities, Modal application and reviewed PR head are in manifest.json. Every complete scorer verdict, ranked time and full energy report was independently reconstructed in verification.json. Copied evidence hashes are in evidence-sha256.json.
The author’s three energy runs and initial no-energy diagnostic below are retained separately and independently rescored in author-record-audit.json. They are not mixed into the reviewer median. The earlier K5-only source failed the previous independent set; all three of those outcomes are retained in historical-review. This repair changes the readout and therefore starts a new three-run qualification set. No failed run was replaced or rerun for either source version.
The earlier K5-only submission failed one independent review run at2.72909% error. These historical local records remain in the parent directory; they do not establish qualification. Independent review.
This replacement keeps E8 width256, three RMSNorm/ReLU hidden layers and1250updates, and adds K5/query-graph propagation (alpha0.9, three steps). It uses CUDA graphs with native kernels and cuBLAS; it is not a whole-model megakernel. Source-only archive, no fitted data.
| Run | MNIST error | Ranked ms | Above-idle mJ |
|---|---|---|---|
| 1 | 2.608182% | 415.527 | 91822.431 |
| 2 | 2.633636% | 422.589 | 91854.397 |
| 3 | 2.530000% | 422.287 | 99702.583 |
All three author-supplied initial energy-enabled canonical Modal runs passed. Median 422.287ms, median energy 91854.397mJ, aggregate error 2.590606%. Initial no-energy diagnostic also retained. No failing run was replaced.
Source SHA256: c6082659a49a6bd942077e359db42f28016626bb49303d68032467710d765a7e; 17973bytes. Independently recomputed all scorer verdicts/ranked times and checked energy problems. The leaderboard uses the independent reviewer medians above; this table preserves the author-supplied results.
Reproduce from mnist-a100:
python run_modal.py submissions/native-mlp-d3-20261007/graph1250-repair/submission.py:classify --difficulty 3 --runs 3 --json out/d3-graph1250