How much of adaptation is organizational: 7% and 10%¶
Superseded on 2026-09-23. Qt's 10% includes its lint bot (bot audit); both shares are corpus v1, organization against organization, and are not re-measured on the current design.
The decomposition the study exists to make, averaged over the three seeds of the windowed run:
| base model | adapter trained elsewhere | adapter trained at home | |
|---|---|---|---|
| OpenStack's refinements | 0.048 | 0.291 | 0.308 |
| Qt's refinements | 0.039 | 0.324 | 0.355 |
Adaptation is worth +0.260 and +0.316 exact match; being adapted on the right organization is worth +0.017 and +0.031 of that, 7% and 10%. Almost everything an adapter learns about writing this corpus's refinements transfers across the organizational boundary: the task is shared, the house style is a tenth of it at most.
That ratio is the finding a reader should carry away, and it is what makes the measurement hard: the organizational term is a tenth of the effect the apparatus can see easily, which is why the design needs the seeds, the crossed interval and the sensitivity analysis to say anything honest about it. It also frames RQ2, where the same small term is enough to identify a source from a weight update.