RQ1's central finding has not been measured before¶

Superseded on 2026-09-23. The share was measured on corpus v1 with Qt's lint bot in the data; the bot audit found nearly half of Qt's organizational effect was the bot. The share on the cleaned corpus is not re-measured in this entry.

A novelty check on the result the windowed run produced: that adaptation to code-review refinement is large and almost entirely task-general, with the organization-specific share about a thirteenth of it (0.048 to 0.306 exact match, of which 2 points separate matched from mismatched).

The closest work fine-tunes for exactly this task and metric and does not make the decomposition. Begolli, Aksoy and Neider (arXiv:2507.19271, v2 October 2025, read in full) fine-tune for code refinement scored by exact match and BLEU, on five industrial C# repositories from one partner plus the CodeReviewer benchmark. They do not compare within-project against cross-project performance, and do not separate task-general gains from organization-specific ones; they attribute what they observe to general C# patterns and domain shift. LLaMA-Reviewer (arXiv:2308.11148) likewise reports the gain from parameter-efficient tuning without decomposing it.

So the matched-versus-mismatched design does something the literature has not: it splits a fine-tuning gain into the part any organization's data would have taught and the part only that organization's could. That is the contribution to state first in the Stage 1 introduction, and it holds whether RQ1's gate passes or fails, because the decomposition is the measurement and the gate is only one reading of it.