H1's AI-trailer and backport sensitivities implemented, read on the development pilots¶

Two of the readings registered beside the pass rule (registered decisions, 2026-10-02) recompute an H1 cell without a set of examples. scripts/sensitivity_ids.py lists them per organization and window from the corpus the partitions were built from (corpus-v3, deduplicated and split as the study loads it): examples whose (project, Change-Id) ai_trailers.py flagged, and examples whose (project, Change-Id) sits only on backport branches (sphragis.corpus.backports, now the one rule backport_share.py reads too). partition_pilot.py --without NAME=FILE reads the cell again without them, on the same runs, levels, seed and draws, beside the cell and binding nothing, and both gates carry the readings. On the development pilots (partition-pilot-<org>-sensitivities.json, 22 runs each; the registered cell reproduces the committed pilots exactly), generated from the artifacts:

The listed sets match the earlier counts (45 AI-flagged OpenStack development examples, 0 and 2 backport-only); 5 of OpenStack's 45 and both of Wikimedia's backports are not among the examples the partitions score, which per-half deduplication and the projects the split excludes account for. Removing OpenStack's AI-flagged examples moves its cell by less than a thousandth. The third registered sensitivity, reviewer exposure, is defined on the organization contrast (sibling minus foreign), which the partition design's H1 does not read; it is implemented with the H2 read, if Qt and Chromium are admitted.

After two review rounds: both gates refuse a test read that lacks either registered sensitivity; a listing records every example it was made from, and the pilot refuses one that does not cover the cell's examples; removed counts examples that left the cell. The searches are bounded by merge date and the windows by creation date, so sensitivity_ids.py also counts examples the searches never reached (changes merged on or after the search bound, or never): none, on either organization's development window. What runs today is the development window: the corpus loader refuses the sealed test window, as the runner does. ai_trailers.py --through-test searches through the test window to the day of the fetch and tallies it, but only once every seal records acceptance, so the test window's listing follows the unsealing.