Why one organization shows the effect and the other does not¶
Superseded on 2026-09-23. A post hoc subgroup reading on corpus v1; Qt's side of it rested largely on its lint bot (bot audit). Not a finding.
The dev-window contrast broken down by the project each held-out change belongs to, pooled over the three seeds. Exploratory, on the dev window, with small per-project samples and no multiplicity correction: a description of where the organizational average comes from, not a test.
| Qt, by project | contrast | examples | OpenStack, by project | contrast | examples | |
|---|---|---|---|---|---|---|
| qt-creator | +0.021 | 291 | starlingx/docs | +0.041 | 138 | |
| qtbase | +0.030 | 135 | starlingx/test | -0.050 | 47 | |
| qtdeclarative | +0.017 | 98 | starlingx/distcloud | -0.078 | 34 | |
| qtdoc | +0.012 | 81 | openstack/kayobe | +0.027 | 25 | |
| qtopenapi | +0.023 | 57 | starlingx/update | +0.042 | 24 | |
| qtmultimedia | +0.000 | 56 | openstack/ironic | -0.030 | 22 |
Qt's effect is spread across its projects; every project with more than fifty examples is at or above zero. OpenStack's projects disagree in sign and average to nothing.
This is the same picture RQ2's client updates gave, from the generative side. OpenStack's Gerrit hosts a federation of loosely coupled projects, StarlingX and Zuul among them, whose conventions need not agree; Qt's projects are one house. So "organization" is a useful boundary exactly when the organization's projects share conventions, and the gate's mixed verdict is not two noisy measurements of one quantity but two different situations. RQ1's registered claim, that organizations leave a learnable fingerprint, is therefore too strong as stated for a federation and about right for a house, which is a finding rather than a failure, and it is what the conjunctive rule was built to expose.