Decided: H1 stays intersection-union over the admitted Gerrit organizations; the GitHub organizations are a registered replication on a second platform¶

AJ delegated the route spec's two open questions (highest-quality research, decided from current practice). Both are decided before any GitHub organization has a corpus beyond its pilot month and before any test window is fetched. This replaces the 2026-10-01 rule that listed the GitHub organizations as candidates behind Qt and Chromium.

1. How H1 reads over more organizations: the intersection-union rule stays. H1 asks whether learned style transfers within the evaluated projects in every admitted organization, which is the claim a perimeter drawn around an organization needs; an intersection-union test makes it at the registered level with no adjustment across cells (Berger, Technometrics 1982). Each added cell raises every cell's power target, 0.95^(1/k): 0.9747 at two cells, 0.9830 at three, 0.9873 at four, and OpenStack's bound at 97.5% moves from +0.0230 at one cell to +0.0248 at two and +0.0261 at three (partition-sensitivity-openstack.json, by_cells). A per-organization reading with a heterogeneity estimate was the alternative. With two to four confirmatory cells the between- organization variance is not estimable with useful precision: the Cochrane Handbook (6.5, November 2024, chapter 10) uses the Hartung-Knapp-Sidik-Jonkman interval only above two studies, warns it is overly wide with very few, recommends prediction intervals from about five, and advises against meta-regression below about ten studies. So the pass rule is unchanged, and two readings beside it answer the other question, neither binding a verdict: the partial conjunction "at least r of k" (registered 2026-10-02), and a random-effects summary over every organization read, Gerrit and GitHub (REML, the HKSJ interval from three organizations, a prediction interval from five), with each platform's subgroup summary reported descriptively and no moderator test. If Qt and Chromium are both admitted, H1 has four cells, and every cell's simulation is rerun with --cells 4 before the seal; the artifacts hold one to three today.

2. The GitHub organizations: a registered replication, not H1 candidates. Two reasons, either sufficient. Platform. On GitHub, review before merging depends on each project's settings and is often optional, where Gerrit's is enforced before a change lands, and GitHub pull requests carry fewer review comments per file (Sun, Wu, Assunção and Stolee, arXiv:2603.15935, March 2026, comparing test-code review on both). Inside an intersection-union H1, a GitHub null would fail the hypothesis for a reason the design does not study. Power. This study adopted the Nature registered-report standard of a priori power 0.95 for every proposed hypothesis test (2026-09-30), so each confirmatory cell needs its simulated bound at Stage 1. Each GitHub organization has one built pilot month. A corpus, its admissible list, a development pilot (22 runs at OpenStack's measured 1.44 GH200-hours each, 2026-09-29) and a simulation (24 and 32 hours on 32 cores for Wikimedia and OpenStack, SPORC jobs 21794712 and 21794711) cannot be ready for every candidate by 2026-11-20, and a cell without its bound at Stage 1 cannot be confirmatory. Chromium's corpus deadline (2026-10-23) is the precedent for admitting only what is ready. Adding the organizations by amendment after in-principle acceptance, as a secondary hypothesis, was the other option; it is a protocol change that needs the editors' approval, so the family is registered now instead.

The family, as registered. Membership is fixed at Stage 1: the organizations, in the order Apache, LLVM, .NET, Grafana (sizing table above), whose pilot, train and development windows are collected, frozen and split-checked at N = 1,850 by 2026-11-20. The Stage 1 report names them; one not ready by then is reported as not collected, and no organization joins afterwards, so the family and the summary across organizations cannot be chosen after any result. Each member is read on H1 only, with the Gerrit cells' windows, rules and estimand, on the two-sided 97.5% interval (one-sided 0.0125, the stricter Holm step); its K comes from its own development pilot and its bound from its own simulation, both fixed before its test window is fetched, so no bound is chosen after the data. The family's reported outcome is the partial conjunction r over its members at one-sided 0.0125, with every cell's verdict; it binds no verdict on H1 or H2. .NET and Grafana are RQ2's federation members whether or not they enter the family. OpenJDK and HashiCorp stay borderline and are not collected unless RQ2's sizing needs them.