The qt-creator fingerprint is a property of the pair, not the project¶
The two remaining pairs among qt-creator, qtbase and qtdeclarative, same protocol as job 148203, one seed each. qtdeclarative is the smallest of the three, so both new pairs equalized at 788 training examples per side, against 1,517 for the first pair.
| pair (a vs b) | train per side | a: matched minus mismatched | b: matched minus mismatched |
|---|---|---|---|
| qt-creator vs qtbase | 1,517 | +0.079 [+0.041, +0.118] | -0.019 [-0.046, +0.008] |
| qt-creator vs qtdeclarative | 788 | +0.005 [-0.024, +0.032] | +0.042 [+0.005, +0.074] |
| qtbase vs qtdeclarative | 788 | -0.024 [-0.060, +0.008] | +0.017 [-0.025, +0.059] |
qt-creator's home advantage does not replicate. Against qtdeclarative, qtdeclarative's adapter scores 0.323 on qt-creator's refinements against qt-creator's own 0.328. So the earlier entry's claim that qt-creator "carries something learnable that is specific to it" holds relative to qtbase only. Two of six directional contrasts exclude zero, both on the supporting side, in different projects.
qtbase never wins at home. Its adapter loses on its own refinements to both neighbours' (0.241 against 0.259, 0.243 against 0.267). qtbase teaches an adapter less than the other two do, at home and away, which reads as heterogeneous or harder training data rather than an absence of habits.
The pairs differ in training size as well as in projects. The first pair trained on twice the
data. Job 148408 reruns qt-creator vs qtbase cut to 788 per side (TRAIN_SIZE=788, a nested
subsample of the 1,517), which separates the two explanations for qt-creator's +0.079.
What survives for Plan D. The separability probe still reads the codebase and not the organization. The generative contrast between projects is real in two directions of six and not organized by project, so "the codebase is the unit" is not yet supported by the contrast itself.