The Qt result was the null's size, not Qt¶
*-cpp-rebuilt-c128.json. The ten-project AOSP rebuild gives seven C++ projects that can fill a
client at 128 examples, against Qt's six, so the project-level permutation enumerates
C(13, 7) = 1,716 distinct groupings where every earlier run had 84. Same training length, same
initialization, a larger and better-balanced project set.
| within C++, 128 examples a client | 3 AOSP projects, 84 groupings | 7 AOSP projects, 1,716 groupings |
|---|---|---|
| AOSP detector, rounds of 16, two clients | AUC 0.940, TPR 0.588 | AUC 0.972, TPR 0.787 |
| Qt detector, rounds of 16, two clients | not measured at 16 | AUC 0.928, TPR 0.564 |
| AOSP project permutation | p 0.012 at the floor, all seeds | p 0.0006 at the floor, all seeds |
| Qt project permutation | p 0.018 median (0.012 to 0.036) | p 0.070 median (0.028 to 0.143) |
| organization within C++, nearest class | 0.765 against a 0.735 majority | 0.646 against a 0.521 majority, p 0.030 |
| organization beyond project, C++ | 0.676, p 0.19 | 0.583, p 0.30 |
AOSP's result strengthens by a factor of twenty in the null it survives. A p of 0.0006 is the floor of 1,716 groupings: no relabeling of which projects belong to which organization scores as well as the true one, at any of four round-draw seeds. The detector reads 0.787 at one false alarm in a hundred.
Qt's does not survive. At 84 groupings and this training length Qt sat near the floor, and the second packing had already softened that to a median 0.042. With thirteen projects and a null twenty times larger it is 0.070, and two of four seeds are above 0.10. The earlier reading came from the small null and the three-project AOSP set it was built against, not from Qt.
What this costs and what it buys. The claim that both organizations become detectable at 128 examples a client is withdrawn. What replaces it is stronger where it holds: one organization is detectable from projects other than the target's, at the floor of a null that enumerates every way its projects could have been relabelled, and the detectability rises with local training length. The asymmetry between the two organizations is now the finding rather than an inconvenience, and it is the same asymmetry the RQ1 side sees, where AOSP behaves as one unit and Qt does not.
The classifier is unchanged by the rebuild: 0.583 beyond project at p 0.30, with classes nearly balanced at 23 against 25, so the operating-point distinction that the rest of RQ2 rests on holds.