Skip to content

Registered before Stage 1: the confirmatory cells are a rule over the admitted organizations

The granularity design had two fixed cell sets, with and without Chromium, and both put Qt in H1. Since 2026-09-26 an organization enters a confirmatory cell only from a host permitting automated access in writing, so Qt, Chromium and Wikimedia are each admitted or not independently, and two fixed sets no longer cover the cases. The cells are now built by one rule from the admitted set (design() in sphragis/experiment/decomposition.py), fixed here before any Wikimedia contrast exists:

  • OpenStack is always admitted; Wikimedia if its split meets the criteria Chromium was held to, frozen by 2026-10-23; Qt if permission arrives before 2026-11-20; Chromium if permission arrives and its corpus is frozen by 2026-10-23 and meets the criteria.
  • H1's confirmatory cells are every admitted organization. H2's are Qt against Chromium and back, only when both are admitted, since that remains the only pair writing one language.
  • An organization without a confirmatory H2 cell is read against the first other admitted organization in the order OpenStack, Wikimedia, Qt, Chromium, as exploratory.
  • The gate takes the admitted set, never a list of cells, and refuses an organization outside the four or a set without OpenStack. A confirmatory cell still needs its registered detectable effect, so Wikimedia's cells cannot be read until the sensitivity analysis is rerun on its corpus.

As of today neither permission request has an answer (QTQAINFRA-8064 has no comment; the infra-dev thread has only an automatic reply), and Wikimedia's build has completed two of its thirteen months.

Added after an independent review, still before any Wikimedia contrast.

  • H1's "why" had been measured for OpenStack and Qt and only assumed for a new organization. The split rule puts the two largest projects in opposite halves, and Wikimedia's two largest (operations/puppet and mediawiki/core) write different languages. So the split criteria gain one: the two halves' file-type mix over training hunks may differ by no more than OpenStack's halves do, by total variation distance over file suffixes.
  • Marginal power stays 0.928 per cell, so the joint power of the intersection over admitted H1 cells is about 0.93, 0.86, 0.80 and 0.74 for one to four cells, stated for the count admitted.
  • Admission is fixed at Stage 1 submission. An admitted organization whose test window cannot be collected after acceptance is reported as not collected and as a deviation, and each hypothesis is read over its remaining cells at the registered Holm levels.