The repeated-partition design on the development window: it stops at 16 runs, and H1 is inconclusive at +0.007¶

Twenty-two partition_run jobs on corpus-v3/openstack, the first 22 admissible partitions at training seeds 1 to 22 and N = 1,850 (jobs 208523 to 208530 pinned at 8ec4e12, 208535 to 208548 at 205ba2f; the two differ only by scripts/partition_pilot.py, which no job runs). Every outcome-neutral check passed and the apparatus held in all 22, and every run scored the same 501 examples: dedup-first left nothing to drop.

Stopping. The rule, applied in order from K = 8, first holds at K = 16 (variance of the mean 1.18e-5 against the critical value 1.30e-5; 1.34e-5 at K = 15). Runs 17 to 22 were computed in the same batch and stay out of the cell. The cut is made in code (partitions.first_stop, which the cell applies to every run it is given and records with its trace). That runs computed in a batch past the first stop are left out was written into the spec as an amendment after both readings were in view (16 runs +0.0072, 22 runs +0.0051, both inconclusive); it is the reading Ritzwoller and Romano's sequential rule defines, and the verdict is the same either way. Per-run H1 estimates over the 16 runs range from -0.0174 to +0.0339.

The cell (partition-pilot-openstack-stopping.json, 16 runs): H1 +0.0072 [-0.0061, +0.0210] at 95% and [-0.0080, +0.0229] at 97.5%, inconclusive. At all 22 runs: +0.0051 [-0.0073, +0.0175]. Read beside the single-partition pilots (v2 +0.0134, v3 +0.0284, both inconclusive or passing on one draw of the partition), the half-split effect in OpenStack's development window is small, and the single-partition v3 pass is not reproduced once the partition is averaged over. A censored pilot, not a test-window estimate; the registered bound for the bounded reading comes from the simulation.