Git audit of two repositories: AI trailer found in only 1 of 39 commits that touch agent notes

Publishers and grant funders increasingly require authors to disclose how AI was involved in their research. For computational work, the answer is usually sought in the repository. Koutian Wu, Zesen Huang and Hejia Geng tested what a repository can actually answer. In their paper "What Git Does Not Record," they analyzed two repositories by the same author in the same field, with 466 and 103 commits. The most telling number: in Project B, 39 commits edit the file where the agent keeps its working notes, yet the Co-Authored-By trailer appears in only one of them, or 2.6%.

Two traces that read differently

Each repository holds its own type of "agentic residue." The first contains persistent rules files, a prescriptive trace. The second contains the notebook of a working agent, a narrative trace.

A rule records a conclusion and erases the history it grew out of. That is why 85.6% of the lines that ever entered the prescriptive trace are visible in the current revision. A notebook records episodes, and only 21.8% of its lines survive in the current version. To an auditor looking at the latest snapshot, the rules look complete and the notebook looks almost empty.

By number of revisions, the gap is much smaller. In the prescriptive trace, 15 of 41 revisions leave no visible lines; in the narrative trace, 18 of 41 do. So the "visible share of lines" and the "visible share of edits" are different metrics, and the choice between them changes the conclusion.

The trailer auditors rely on

Automated audits usually look for a Co-Authored-By field in the history. The authors show that the field is incomplete and unstable:

• Project A has three different AI trailer keys, and Co-Authored-By is only one of them. It covers 66 of the 121 commits that contain any AI marker.
• In Project B, the trailer's sensitivity is low. Measured against an explicit in-repository proxy (commits that change the agent's notes file), the trailer is present in 1 of 39 commits, or 2.6%.
• The share of commits with the trailer in Project A cannot be treated as stable over time. It rises from 2.7% to 17.7% at a boundary the authors did not set in advance.

The authors draw a cautious conclusion: channels inside the repository disagree with each other, and none of these shares can be read as a share of AI involvement. They do not isolate the cause of the divergence and say so explicitly. It could be a change in the author's habits, a change of tool, or a change in agent settings.

What the traces answer, and where they are confidently wrong

Between commit fields and artifact files, the authors build a table: which claims about a work's provenance a trace supports, which it cannot support, and which it answers confidently but wrongly. The last category is the most dangerous. A clean rules file looks like evidence of discipline, even though it only stores the outcome and hides the path. A missing trailer looks like the absence of AI, even though the agent's notes in the same commit say otherwise.

On this basis, the authors built an auditable evidence ledger. Each claim in it is tied to a specific source and to the limitations of that source. From the analysis, the authors derive requirements for systems that should record a study's provenance at the time the work is done, rather than reconstruct it from history after the fact.

Alongside the paper, they released the git-provenance-auditor tool on GitHub (repository ktwu01/git-provenance-auditor, currently one star). The preprint is published as arXiv 2610.28655.

What teams can do right now

The sample is small: two repositories by one author. It will not yield universal percentages, but the practical takeaways are clear.

• Do not rely on a single trailer. If a journal's policy requires disclosure, cross-check Co-Authored-By against other keys and against the agent's notes files.
• Decide in advance what to keep. An agent notebook that gets overwritten loses almost four-fifths of its lines in the current version. If the episodes matter, commit them as separate artifacts.
• Do not publish an "AI involvement share" based on the share of commits with a trailer. The jump from 2.7% to 17.7% in a single repository shows how much the metric depends on habits and settings.
• Record provenance at the time of the work. Who did what should be logged through the agent's tooling or a commit hook, not reconstructed from history.

For publishers and grant funders, the conclusion is stark: a disclosure form that asks for a repository link will get data that looks convincing and may be wrong.