Wikimedia's simulation at the lower-bound window: two-cell bounds +0.0268 and +0.0246¶
partition-sensitivity-wikimedia.json, SPORC job 21794712 (CPU only, pinned at 8c87f28 with the
arguments of OpenStack's rerun), in place of TIGRIS job 211288, which was cancelled when the TIGRIS
queue slipped. Its placebo input is read from the cluster data root; that file, redacted, is
byte-identical to the committed rq1-placebo-wikimedia.json. 1,000 studies per power step and
4,000 null studies, K = 16, N = 1,850, the 2,848-change lower-bound window, and sigma_run
calibrated on the pilot's 429 changes:
| spread point | spread | sigma_run | FP 97.5% | FP 95% | detects 97.5% / 95%, two cells | null bounded (97.5%, two cells) | reproducibility failure |
|---|---|---|---|---|---|---|---|
| pilot 90% lower | 0.0094 | 0.0080 | 0.0145 | 0.0255 | +0.0238 / +0.0223 | 0.987 | 0.000 |
| pilot estimate | 0.0112 | 0.0100 | 0.01425 | 0.027 | +0.0246 / +0.0230 | 0.988 | 0.004 |
| sizing bound (90% upper) | 0.0141 | 0.0132 | 0.015 | 0.029 | +0.0268 / +0.0246 | 0.990 | 0.029 |
| pilot 99% upper | 0.0172 | 0.0165 | 0.016 | 0.0285 | +0.0289 / +0.0276 | 0.989 | 0.088 |
The false-positive rate sits slightly above nominal at every point, by at most about two Monte Carlo errors (0.0018 at 0.0125, 0.0025 at 0.025, over 4,000 null studies). Reproducibility holds at the spread K was sized on (0.029 against beta = 0.05) and fails past it (0.088 at the 99% bound), as sizing at a 90% bound intends.
Read under the registered rule (bounds are the detectable effects at the sizing bound, at the
cell count of the admitted set): with OpenStack and Wikimedia admitted, Wikimedia's bounds are
+0.0268 at 97.5% and +0.0246 at 95%, each cell at power 0.9747. The one- and three-cell bounds
are in the artifact (by_cells) for an admitted set of another size; whether H1 spans more cells
is an open decision. Because the window is a lower bound, these bounds are conservative. The
Stage 1 figures move once OpenStack's rerun (job 21794711) lands.