sutro-problems

D5 kernel-taught Ladder: width 768, 480 fused Adam updates

Independent reviewer reproduction — October 19, 2026, later pass

The first frozen reviewer group passed 3/3 fresh canonical sandboxed Modal A100-80GB runs. Median ranked time is 492.842637 ms/call; median above-idle energy is 77,089.864599 mJ/call. Pooled MNIST error is 1.792424% (324,085/330,000 correct). Each run independently passes every accuracy, holdout and timing gate; pooled accuracy is descriptive.

Reviewer run MNIST error Margin, labels Ranked ms Above-idle mJ
1 1.807273% +102 492.842637 71983.093171
2 1.697273% +223 492.085754 77089.864599
3 1.872727% +30 495.066059 80849.184445

The submitted source is unchanged, SHA-256 e8d330a1a81b139f5b89290325b6f7a04123f47dc3c84af354b141f96a059b53; scorer and stock runner are unchanged. All 45 timed calls completed. No outcome was replaced or retried. Raw records and executable verifier, reconstruction and source review are retained. Optional energy is assessed separately from score qualification.

This is @sjbaebae’s budget/kernel optimization of @spartanhaden’s kernel-taught Ladder. Credits to Haden, @SethTS/Rasmus et al., @islamborghini, @likelytobelaura and @sigkillme0 remain in the submission. Teacher fitting, 480 student updates and both predictions execute every call; the label-symmetric guard selects already-computed predictions. Earlier 510-step and failed fast450 groups remain separate from this exact-source group. Development arithmetic reconstructs; unpublished variant source/history claims are not independently authenticated. The author manifest’s pending wording is historical.

The leaderboard uses these independent medians; original author results below remain separate. Published instructions require three official runs and their median; this reviewer reproduction supplies evidence and does not introduce a permanent source ban or an additional published submission rule.

Using Python 3.13, from this submission directory: python evidence/reviewer-20261019-pass3/verify_records.py evidence/reviewer-20261019-pass3 submission.py --difficulty 5 --function ladder_kernel --output /tmp/pr135-verification.json. This audits saved records without executing the learner.

Original author report (as submitted)

Pending-review and record claims below describe the author submission; the dated reviewer section gives accepted status.

This derives directly from @spartanhaden’s kernel-taught Ladder D5, retaining its kernel teacher, reconstruction, tangent jitter, hard-query sampling, fused Triton combinator backward and guard. Original credits remain: @SethTS/Rasmus et al. for Ladder, @islamborghini for kernel-PCG and batch/step work, @likelytobelaura for tangent jitter, and @sigkillme0 for grouped graph replays. This is an architecture/learning-budget/optimizer-implementation change, not a new learning algorithm.

Current frozen candidate

The student changes from 60–1024–512–256–10 to 60–768–512–256–10, and Adam updates decrease from 600 to 480. Batch remains 2500; LR is .01378125, approximately retaining the integrated LR budget. The original constant-first-two-thirds, linear-tail schedule and Adam betas (.8, .999) remain. Adam uses the native fused implementation rather than compiled foreach, still through the compiled optimizer wrapper. Floating-point updates may differ; bitwise equivalence is not claimed. Grouping remains 30 steps per graph replay, so exactly 480 updates execute.

The existing backward kernel receives a width-768 configuration (64,8,4,2,32); its body is unchanged. The teacher’s two metric updates and pseudo-label refit, BN/ReLU Ladder, Gaussian noise .25, tangent jitter, reconstruction and guard are unchanged. Neural activations are BF16; master parameters and Adam states remain FP32. There is one student and no final neighbor lookup. Test features are used without labels for teacher pseudo-labelling, the training pool and reconstruction, as in the parent.

All metric/teacher fitting, student reset/training and prediction occur in each measured warm call. Input-dependent learned state is reset before reuse; only code, graphs and scratch persist. Fused Adam alone was not faster in development: the retained candidate couples it with fewer updates. The complete-call gain is not attributed solely to optimizer fusion.

Three independent author canonical sandbox evaluations

Run Ranked ms MNIST error Above-idle energy mJ
1 491.963506 1.790000% 83887.252169
2 492.013229 1.726364% 84000.853066
3 492.937883 1.810000% 86014.630982

All three pass the unchanged canonical sandbox judge, including all four foreign hold-out calls in every run. Median ranked latency is 492.013229 ms, median above-idle energy 84000.853066 mJ, and pooled MNIST error 1.775455% (324141/330000 correct). Each of the three ranked times is below 500 ms. Three distinct A100-SXM4 GPU UUIDs were recorded. All original outcomes are included; there were no agent-initiated qualification replacements, and no infrastructure preemption is recorded in the execution log.

The listed parent reviewer medians are 633.529424 ms / 107885.521303 mJ. This author group is approximately 22.3% faster and 22.1% lower energy; pooled error is slightly higher (1.775455% versus 1.75%). These are different data draws and GPU hosts, not a paired comparison. Independent reviewer replication is outstanding. The leaderboard is not edited and no accepted record is claimed.

Evidence and reproduction

Current exact source: 20436 bytes, SHA256 e8d330a1a81b139f5b89290325b6f7a04123f47dc3c84af354b141f96a059b53. Packaging matches the development program by AST except updated module prose. Manifest, all current raw records and verification are included. Saved-record verification independently rebuilds the judge, score, summary, source identity and every energy window; all checks pass. Scorer and runner hashes match the tracked versions. This audit is not a second execution group.

python run_modal.py submissions/ladder-width768-d5-20261019/submission.py:ladder_kernel --difficulty 5 --runs 3
cd submissions/ladder-width768-d5-20261019
python verify_records.py evidence/official submission.py --difficulty 5 --function ladder_kernel --record-filename classical-width768-fused480-d5-20261019.py --output evidence/official/verification.json

Prior candidates and development

This PR initially contained a distinct 510-update compiled-foreach candidate with a qualifying 507.518126 ms median and 1.780000% pooled error. Its exact source, manifest, all three passing raw records and re-audit are retained in evidence/prior-width768-s510. No records were pooled across candidates to choose a median. Its original commit remains in PR history. Re-audit it with:

python verify_records.py evidence/prior-width768-s510 evidence/prior-width768-s510/submission.py --difficulty 5 --function ladder_kernel --record-filename classical-width768-s510-d5-20261019.py --output evidence/prior-width768-s510/verification.json

An earlier, different 1024-wide/450-update candidate failed one of three canonical runs by eight predictions. Its complete original group remains separately in evidence/prior-fast450; it has no qualifying median. No failed outcome was replaced by a passing run.

Development measurements are warm complete calls, not canonical qualification. Eleven fresh paired inputs with three stochastic retrainings per input measured compiled-foreach/510 at 1.823939% / 509.653650 ms, native-fused/480 at 1.842424% / 494.013724 ms, and native-fused/510 at 1.829697% / 525.629177 ms. Fewer updates reduced runtime with slightly worse development accuracy; the exact final-source canonical group above independently qualifies. Full development records are included. Other batch/readout summaries are exploratory and are not part of this classifier.