GPU-only one-seed iteration

Completed results and timing are in b1-seed42-results.md. The commands below record the completed 2,048-draw execution. New evaluations default to 256. The overnight commands below resume the original pinned 2,048-draw plans; they must not be used for a new 256-draw comparison.

2026-09-17: the user cancelled the ten-seed daytime execution and its final report, retaining the plans for an overnight run. The proposed CPU screen was withdrawn before submission; its launcher, instructions and unused experiment folder were removed. No CPU fits ran. The decisive-report inference default is CUDA as well; GPU allocation failure must not trigger a CPU fallback.

Run seed42 only for all A/B/C candidates, with the full 100-epoch cap, patience30, 128 training members, 256 validation members and 2,048 evaluation members. Keep all three folds and select-then-refit. Reuse A's verified complete seed42. B and C run on separate L40 GPUs without the ten-seed contention. The rough 25–40 minute estimate comes from A's prior ~24-minute full seed runs, not a guarantee.

The dispatcher accepts --seeds 42 to restrict this invocation without changing the ten-seed plans. manager status/rank --seeds 42 reports this subset. Omit the seed filter when resuming the full plans overnight. Interrupted incomplete attempts remain available for audit; the manager starts fresh attempts for incomplete seeds. Completed seeds remain reusable.

This is a paired engineering screen, not a ten-seed conclusion. There is no estimable across-seed uncertainty with one seed, and no ten-distribution mixture. Use the shared natural/stress forecast score and calibration to guide iteration; the original uncertainty and decision-rule report waits for the overnight run.

One-seed report command after the two new fits finish:

.venv/bin/python -m tapestry.models.manager decisive -e B1-onlymask-refit \
  --seeds 42 --device cuda --report-output data/experiments/B1-seed42-report

The same dependency report launcher takes an optional seed (scripts/b1_decisive.sbatch 42). It uses its separately pinned reporting source. The report includes all three pairwise differences, corrected stress scores, target/season/geography calibration, C's recent mechanisms, and conditional temporal intervals. It does not manufacture seed intervals or a ten-member-model mixture from a single fitted seed.

Submitted jobs and exact commands

The full daytime jobs 1487036, 1487037, 1487038 and pending full report 1490284 were cancelled at the user's request. A seed42 is reused from its complete attempt; B/C seed42 restart as attempt002. New GPU jobs: B=1491695, C=1491696. The seed42 report is job1491906. The training source snapshot is 2f10c91; the independent report snapshot is 2b50923. Notifications remain enabled.

for e in B1-direct-finalflag B1-joint-aux025; do
  .venv/bin/python -m tapestry.models.manager plan -e "$e" \
    --suite "$e" --seeds 42 --eval-members 2048 --retrospective --device cuda
  LANES=1 GPUS=1 sbatch --job-name="$e-s42" --array=0 \
    scripts/jlessler.sbatch "$e" --seeds 42
done
for e in B1-onlymask-refit B1-direct-finalflag B1-joint-aux025; do
  .venv/bin/python -m tapestry.models.manager status -e "$e" --seeds 42
  .venv/bin/python -m tapestry.models.manager rank -e "$e" --seeds 42
done
sbatch --job-name=B1-seed42-report --dependency=afterok:1491695:1491696 \
  --kill-on-invalid-dep=yes scripts/b1_decisive.sbatch 42

For the overnight run, the original ten-seed plans remain intact. Once the seed42 jobs have ended, omit the seed restriction to resume unfinished seeds:

LANES=7 GPUS=1 sbatch --job-name=B1-onlymask-refit --array=0 scripts/jlessler.sbatch B1-onlymask-refit
LANES=10 GPUS=1 sbatch --job-name=B1-direct-finalflag --array=0 scripts/jlessler.sbatch B1-direct-finalflag
LANES=5 GPUS=2 sbatch --job-name=B1-joint-aux025 --array=0-1 scripts/jlessler.sbatch B1-joint-aux025

Use the three newly returned job IDs in --dependency=afterok:ID_A:ID_B:ID_C when submitting scripts/b1_decisive.sbatch without a seed argument. No overnight restart has been scheduled automatically.