Qt is not a coherent organization in update space, and that is the finding¶

Correcting the entry above, which said the pooled defence curve read Qt below chance because the reference direction was really a language direction, and that holding content fixed would settle it. It did not. Restricted to the 43 C++ clients, Qt's curve still read 0.364 at 10 rounds.

I then supposed the cause was a split mismatch: the aggregate attack splits its reference over projects while the defence curve split randomly over clients, and mirroring the split would fix it. That was also wrong, and the test refuted it. With the project split Qt reads 0.323, slightly worse. The change is kept anyway, because splitting over projects is the principled choice and its sibling detector already did it -- a guard fixed on one path and not on the path beside it.

The direct measurement had the answer, within C++:

Qt's C++ clients are less like each other than they are like AOSP's. Held-out minus outsider alignment with the organization's own reference is +0.045 for AOSP and -0.008 for Qt, the same sign in all six splits. Splitting over projects makes Qt worse because within-Qt coherence is almost entirely a project effect (same project 0.2597, other project 0.2523), and a project split removes exactly that.

So the below-chance reading is not a defect to be repaired. It says a round holding a Qt client looks less like Qt's other projects than a round of AOSP clients does. AOSP's three projects are Android platform C++ under one style guide and one build system; Qt's six span a framework, an IDE and a QML runtime. At 64 examples a client, "Qt" is a name over a federation of unlike codebases and "AOSP" is a house.

This is the same fact every other instrument has been reporting. Qt's project-level permutation is p = 0.238 from the product and from B, against AOSP's 0.012; Qt only reaches the 1/84 floor from A alone, or at 128 examples a client. A near-zero quantity is exactly where two statistics can disagree in sign, and the aggregate detector's +0.613 for Qt rests on a gap of 0.004 on a base of 0.69. The honest reading is that Qt's organizational signal at this training length is not reliably present, not that one instrument is broken.

It also answers the question the RQ1 sharpening was circling: what makes an organization one? An organization is a detectable unit exactly when its projects resemble each other more than they resemble an outsider's. That is a measurable property, it differs between the two organizations in this study, and it is not something a researcher gets to assume by drawing a box around a GitHub org. Norm is not the cue: AOSP's updates average 2.16 and Qt's 2.11 with standard deviations of 0.04 and 0.03, and rescaling every update to unit norm leaves the detector unchanged (0.613 and 0.699 against 0.613 and 0.704).

Qt's defence-curve cells are therefore withheld from interpretation at 64 examples a client, and the curve is worth rerunning against Qt only at a training length where Qt is coherent.