Research inclusion; D4 qualification and author-source attribution remain unresolved. The submitted 18,558-byte source passed 1/3 independent reviewer runs; two complete runs missed D4 by 26 and 20 labels (all 45 timed calls retained). The author’s 35.678700 ms median below belongs to records identifying a different, unsupplied 18,611-byte source (c7b78944…c52d78b1), not the packaged source (1966a800…83aa77). It is historical author evidence, not a verified median for submission.py. No qualifying reviewer-group median or leaderboard promotion is made. See reviewer evidence.
Attribution: four documented teacher helpers follow @spartanhaden’s kernel/Ladder D4; related tangent/Muon/cross-teaching work is credited to @likelytobelaura and joint graph propagation to @sjbaebae. Complete development lineage remains unsettled. Original source and records are preserved.
author measurements: three passing sandboxed a100-80gb runs per source. results.
| source | median ms | pooled error |
|---|---|---|
| default | 35.678700 | 2.160606% |
gpu reproduction from mnist-a100/, with modal configured:
python run_modal.py submissions/neural-d4-20261018/submission.py:classify \
--difficulty 4 --runs 3 --json /tmp/d4-default
offline check from this directory: python3 -B verify_records.py. records omit console logs and gpu identifiers; filenames normalized.