A codebase fingerprint inside one organization¶

Qualified later on 2026-09-18. qt-creator's home advantage holds against qtbase only; against qtdeclarative it is +0.005, covering zero. See "The qt-creator fingerprint is a property of the pair, not the project" below.

datasets/results/projects-qt-creator_qt-creator-qt_qtbase.json. The identical matched-versus- mismatched contrast RQ1 runs, between two projects inside Qt: qt-creator/qt-creator (a) and qt/qtbase (b), holdout by change within the train window, equalized at 1,517 training examples each, one seed. Both are C++, so the language confound that separates OpenStack from Qt is absent.

Gate: mixed. All seven outcome-neutral checks pass.

qt-creator carries something learnable that is specific to it. Its own adapter beats the other project's on its refinements by 7.9 points, interval clear of zero, roughly four times the +0.021 separating OpenStack from Qt. That is not qt-creator's data being easier: it is easier for any adapter, qtbase's included (0.273 there against 0.241 at home), but the contrast compares two adapters on the identical examples, so difficulty lifts both arms and cancels. What remains is what only qt-creator's training data could teach.

qtbase shows no home advantage, and the reason is visible. The qt-creator adapter is the stronger of the two in general and edges ahead even on qtbase, 0.259 against 0.241. A project whose own adapter cannot beat a neighbour's on its home data either has no distinctive conventions an adapter absorbs, or has them below the floor the calibration found. The fingerprint is a property of some projects, not of every project.

Two measurements with no mechanism in common single out the same project. qt-creator is also the outlier on review latency, settling 76% of its example-bearing changes in their creation month against 51% for qtdeclarative. The separability probe, the latency analysis and now the generative contrast each point at the project as the unit where variation lives, and two of them at this project specifically.

Recorded caveats. The split is holdout by change inside one window, not the time split the windowed RQ1 used, and holdout is the easier test because train and evaluation share a period. The fairer comparison is therefore the organization-level pilots, which were holdout by change too and returned intervals straddling zero, though on far fewer changes. One seed. And one project pair: the claim this licenses is that a codebase-level fingerprint exists and can be larger than the organizational one, not that every project has one.