Research inclusion; D5 qualification remains unresolved. The existing author records reconstruct 3/3 (median 308.420616 ms) separately from the initial reviewer group, which passed 1/3. Two complete reviewer runs missed D5 by 116 and 57 labels; all 45 timed calls are retained. There is no qualifying reviewer-group median and no leaderboard promotion. The submitted source and original author records remain unchanged. See reviewer evidence.
Teacher/Ladder attribution: @spartanhaden’s PR #121, earlier kernel-PCG work by @islamborghini, and Ladder work by @SethTS and Rasmus et al. This is a substantial optimizer/schedule/kernel/readout variant; remaining full lineage is not settled here.
author measurements: three passing sandboxed a100-80gb runs per source. results.
| source | median ms | pooled error |
|---|---|---|
| default | 308.420616 | 1.827879% |
gpu reproduction from mnist-a100/, with modal configured:
python run_modal.py submissions/ladder-d5-20261018/submission.py:ladder_kernel \
--difficulty 5 --runs 3 --json /tmp/d5-default
offline check from this directory: python3 -B verify_records.py. records omit console logs and gpu identifiers; filenames normalized.