The altitude below ours has already been shown learnable¶

Dai et al., "MPCoder: Multi-user Personalized Code Generator with Explicit and Implicit Style Representation Learning" (ACL 2024, arXiv:2406.17255), learn per-user coding style: an explicit residual for syntactic conventions, an implicit representation trained with contrastive learning to separate users, and a Coding Style Score over structure, formatting and naming to say whether the generated code resembles the user it was generated for.

So individual style is learnable and separable, by a method built to separate it. That is the altitude below this study's, and it cuts both ways for RQ1. It supports the premise, since a style signal exists somewhere in the hierarchy; and it sharpens the question, because MPCoder needs contrastive training designed to pull users apart, while RQ1 asks whether ordinary fine-tuning on an organization's own refinements picks up that organization's hand as a side effect. A method built to separate sources and a method that happens to absorb one are different claims, and the 7 to 10% organizational share measured here is what the second is worth.

It also leaves the middle of the hierarchy, the project and the organization, where this study and its RQ2 attacks now sit, with the finding that the codebase is the unit that carries.