A new study audits two single-investigator research repositories to examine how well git history captures the involvement of AI coding agents in producing committed code. The researchers find that most lines in a repository built around standing rule files remain traceable at the current revision, while far fewer lines survive from a repository built around a running agent scratchpad. They also show that the “Co-Authored-By” commit trailer captures only a small share of AI-involved commits in one project, meaning the two provenance signals disagree and neither can be treated as a reliable AI-involvement rate. The authors outline which provenance claims these git-based traces can and cannot support, and derive requirements for systems designed to record research provenance as it happens.
