I audited 569 commits across two of my own AI-assisted research repositories and built a small tool to run the same audit on any public repo. The paper is "What Git Does Not Record" (arXiv:2610.28655, NeurIPS 2026 AI-Native Academia Workshop).
The tool, Git Provenance Auditor, scans git history for AI assistance trailers (for example Co-Authored-By lines from Claude Code and Cursor), counts how often agent working files such as agent-notes get revised, and shows how much of the agent's work the log actually records.
The uncomfortable finding: the trailers cover only part of it. In one repository, 66 of 121 commits carrying any AI trailer use the standard Co-Authored-By line. In the other, 39 commits revise the agent-notes file and only 1 carries the trailer.
If you cite commit logs as provenance for AI-assisted work, most of the trail is missing. That matters for audits, for evaluations built on agent-written code, and for anyone reconstructing what an agent actually did after the fact.
Tool: [https://github.com/ktwu01/git-provenance-auditor](https://github.com/ktwu01/git-provenance-auditor)
Paper: [https://arxiv.org/abs/2610.28655](https://arxiv.org/abs/2610.28655)
Feedback welcome, especially from people maintaining agent-written codebases: what do you wish your git history recorded?