OpenStack on corpus v2 at the half size: the seed effect holds five seeds, and the dev-window H1 pilot is inconclusive at +0.013¶

The first GPU runs on corpus v2. The v2 OpenStack corpus was copied to TIGRIS (corpus-v2/openstack, every file's sha256 matched) and the placebo run at the half size under RUN_TAG=v2, five single-seed jobs (206166, 206171 to 206174), each building its halves into its own root so concurrent builds could not overwrite one another. The five roots are byte-identical, checked on the cluster with find . -type f ! -name '*.source.json' ! -name placebo.json -print0 | sort -z | xargs -0 sha256sum | sha256sum in each root: every one gives f49b745cfbde62c0. --corpus-digest records that value when the runs are merged; it is the caller's assertion and the merge does not recompute it. merge otherwise refuses runs whose corpus sources differ, and with the digest compares them with only the root removed, mode, windows and every count still compared. The halves equalize at 2,158 examples, of which 2,155 trained (three refused by the trainer), and are scored on 116 and 113 development changes.

Seed effect (seed-effect-placebo-openstack-v2.json): 0.0116 at five seeds, one-sided 95% upper bound 0.0426 on 4 degrees of freedom. On corpus v1 it was 0.0106 at three seeds with an upper bound of 0.0865. Still above 0.01, so the registered five seeds stand.

Sensitivity at the half size on v2 (decomposition-sensitivity-v2.json, 100 trials a step, 1,000 resamples). Built as the v1 artifact was: seed 1's placebo run as the variance model and the five-seed point estimate of the seed effect. The planned test total, 1,809 changes, is the v1 projection carried over rather than recomputed (the v2 development window holds the same 229 changes as v1); v2's development proportions split it 916 and 893:

The v1 figures were +0.0273 and +0.0257: one bisection step apart, which at 100 trials a step is indistinguishable at this simulation's resolution. Qt is not rerun: it is not admitted.

The decomposition gate on the development window, as a pilot (decomposition-pilot-openstack-v2.json, scripts/decomposition_pilot.py, OpenStack admitted alone, so H1 is read at one level): the half-split contrast is +0.0134 [-0.0091, +0.0350] at 95%, inconclusive. The estimate is above the smallest effect of interest and below both detectable effects, and the interval does not exclude zero. This is a censored development window at 116 and 113 changes, not an estimate of the test-window effect.

merge and configuration moved from scripts/crossed_reread.py into sphragis/experiment/runs.py, which imports no GPU stack, so seed_effect.py no longer needs the model venv and the merge guards are tested.