The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs. Median ranked time is 406.915422 ms/call; median above-idle GPU board energy is 67,812.879400 mJ/call. Pooled MNIST error is 1.749091% (324,228/330,000 correct). Each complete run independently passes all accuracy, holdout and timing gates; pooled accuracy is descriptive.
| Reviewer run | MNIST error | Margin, labels | Ranked ms | Above-idle mJ |
|---|---|---|---|---|
| 1 | 1.790909% | +120 | 406.915422 | 67812.879400 |
| 2 | 1.761818% | +152 | 406.980608 | 67766.748575 |
| 3 | 1.694545% | +226 | 404.012348 | 71178.475965 |
The submitted source is unchanged, SHA-256 98e310b12031555088ee47e6f8170ad001fad1b6f0be28dcead493ff7682f5c3; scorer and stock runner are unchanged. All 45 timed calls completed. Every initial outcome is retained; no maintainer-launched replacement or retry was made. Raw records and executable verifier, reconstruction, decision and source review are retained. Optional energy is assessed separately from score qualification and excludes host/system energy and initial compilation.
This is @spartanhaden’s kernel-taught Ladder at @sjbaebae’s width-768 budget, retaining the @SethTS/Rasmus et al., @islamborghini, @likelytobelaura and @sigkillme0 lineage. Labelled-row top-down input credits the related #127 idea, implemented with current training labels only; query targets remain predictions. The per-call noise bank, teacher change, launch configuration and hidden-space HKNN are stated separately from earlier budget/kernel work. Large development batches, sanitizer claims and unsubmitted variant histories remain author claims unless supported in the retained audits. The separate exploratory D6 study is research and does not qualify this submission.
The leaderboard uses these independent reviewer medians; the original author report below remains separate. Published instructions specify three official runs and their median; this reviewer reproduction supplies evidence, without adding a permanent source exclusion or a new mandatory second-batch submission rule. No D6 qualification is claimed here.
Using Python 3.13 from this submission directory: python evidence/reviewer-20261022/verify_records.py evidence/reviewer-20261022 submission.py --difficulty 5 --function ladder_kernel --output /tmp/pr139-verification.json. This audits saved records without executing the learner.
Present-tense record and pending-review claims below describe the author submission date; the dated reviewer section above gives accepted status.
Entrant: @spartanhaden.
This entry builds on our merged ladder-k2-d5 (#129) at #135’s width of 768, with 450 Adam steps instead of 480, and changes the teacher, noise source, decoder top-down input and readout.
Three sandboxed Modal A100-80GB runs of the unchanged scorer (mnist-a100/1.2.0, stock run_modal.py --difficulty 5 --runs 3) all passed.
| Run | GPU | MNIST error | Margin (labels) | Worst draw | Hold-out | Ranked ms | mJ above idle |
|---|---|---|---|---|---|---|---|
| 1 | A100-SXM4-80GB | 1.692% | +229 | 1.94% | KMNIST 97.41% | 405.073 | 69,649 |
| 2 | A100-SXM4-80GB | 1.720% | +198 | 1.86% | KMNIST 97.30% | 403.997 | 64,250 |
| 3 | A100-SXM4-80GB | 1.720% | +198 | 1.92% | KMNIST 97.37% | 405.940 | 64,777 |
Compared with the current record, @sjbaebae’s ladder-width768-d5 (#135; reviewer median 492.843 ms, 77,090 mJ), this entry is 1.22× faster and uses 16% less energy per call.
submission.py is 20,474 bytes, 6 under the size limit, with SHA-256 98e310b12031555088ee47e6f8170ad001fad1b6f0be28dcead493ff7682f5c3, identical to source_sha256 in all three records (working name x_k5cFdt450.py). It passes check_source, and evidence/official/manifest.json hashes every file.
v + (label − v).detach()); test rows keep the softmax. At 480 steps it gives −0.048 ± 0.002 points for +1 ms (2,112 paired fresh A100 calls). At 450 steps this file is still −0.028 ± 0.002 points better than the 480-step file without it, and 24.5 ms (5.7%) faster.comb_bwd_kernel, with its body unchanged: −10 ms, +0.000 ± 0.008 points.Certified on 2,112 fresh in-process A100 calls: mean error 1.743%, 11-draw split sd 0.058, 0 of 192 splits over the band (worst 1.896%), P(3/3) ≈ 0.990. No call went over the per-draw limit after the guard, which tripped once.
Under compute-sanitizer on a Modal A100, comb_bwd_kernel showed 0 memcheck errors, 0 racecheck hazards and 0 synccheck errors on our 480-step file, whose kernel text, launch configs and shapes are identical.
k5_cFdt480.py, SHA-256 218e3643…, median 429.512 ms) and that file without the label input (k5_cF.py, SHA-256 e6b3ac01…, median 431.666 ms). Our ladder-k2-d5 README lists the earlier sets.Reproduce from mnist-a100/:
python run_modal.py submissions/ladder-label-topdown-450-d5-20261020/submission.py:ladder_kernel --difficulty 5 --runs 3 --json /tmp/d5-verification
Co-developed with Claude.