sutro-problems

Kernel-taught Ladder with a fused Triton combinator backward: difficulty 5

Independent reviewer reproduction — October 19, 2026

The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs, with valid above-idle energy measurements (3 recorded GPU UUIDs). Median ranked time is 633.529424 ms/call and median energy is 107,885.521303 mJ/call; pooled MNIST error is 1.750000% (324,225/330,000 correct).

Reviewer run MNIST error Margin, labels Ranked ms Above-idle mJ
1 1.750000% +165 630.967276 118,481.918772
2 1.710909% +208 633.529424 107,885.521303
3 1.789091% +122 636.385705 106,205.624267

Source is unchanged at SHA-256 cc423b95ed940a5bcd5dc3b08f71502dddbcee3b42cbc34fc5322819378aef2a. The scorer and runner are unchanged. Verification reconstructs every complete canonical score and energy result from the raw records; source review covers per-call fitting, label isolation and cached-state reset. No reviewer outcome was replaced or retried.

This is @spartanhaden’s expanded kernel-taught Ladder variant: the accepted Ladder/kernel teacher and D4 teacher, @SethTS/Rasmus et al.’s Ladder work and @islamborghini’s kernel-PCG remain credited. It also credits @islamborghini’s batch/step tuning, @likelytobelaura’s tangent jitter, and @sigkillme0’s held #127 for grouping steps in graph replays. The branchless guard chooses between already-computed predictions; it does not skip the teacher or any of the 600 student updates. Exact-source author records use working filename a_k2t_600.py and reconstruct separately. Development-call counts, probability estimates and historical kernel chronology below are author assertions, not independently reproduced qualification evidence.

The leaderboard uses the independent reviewer medians above. Author results below are retained separately. The published instructions specify three official Modal runs and their median; this independent reproduction is review evidence, not a newly imposed second-batch submission rule.

Using Python 3.13 as in the canonical runner, recheck saved records from this submission directory with python evidence/reviewer-20261019/verify_records.py evidence/reviewer-20261019 submission.py --difficulty 5 --function ladder_kernel --output /tmp/pr129-verification.json. This reconstructs recorded verdicts without executing the learner.

Original author report (as submitted)

Present-tense record and pending-review claims below describe the submission date; the dated reviewer section above gives the accepted review status.

Entrant: @spartanhaden.

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 1.850% +55 2.05% KMNIST 97.26% 632.089 112,581
2 A100 80GB PCIe 1.708% +211 1.98% KMNIST 97.27% 636.150 110,725
3 A100-SXM4-80GB 1.636% +290 1.89% KMNIST 97.34% 630.646 119,454

Compared with the current record (897.5 ms, 154,289 mJ; @islamborghini’s tuning of our Ladder), this entry is 1.42× faster and uses 27% less energy per call.

submission.py is 20,474 bytes with SHA-256 cc423b95ed940a5bcd5dc3b08f71502dddbcee3b42cbc34fc5322819378aef2a, identical to source_sha256 in all three records (which name it by its working name, a_k2t_600.py). It passes check_source. evidence/official/manifest.json lists the hash of every file.

Method

Certified before the official set on 4,301 fresh in-process A100 calls: mean error 1.739%, P(3/3) ≈ 0.99.

Credits

Notes

Reproduce from mnist-a100/:

python run_modal.py submissions/ladder-k2-d5-20261019/submission.py:ladder_kernel --difficulty 5 --runs 3 --json /tmp/d5-verification