Two AI-assisted research repos leave incomplete Git provenance
Original title: What Git Does Not Record: Provenance Legibility in Two AI-Assisted Research Repositories
Co-Authored-By marked 66 of 121 AI-related commits in one repo and 1 of 39 agent-note commits in the other.
The authors audit two single-investigator repositories in the same field, with 466 and 103 commits, looking for AI contribution traces. A prescriptive trace of standing rules keeps 85.6 percent of its ever-committed lines at the current revision, while a narrative agent scratchpad shows 21.8 percent. By revision, 15 of 41 prescriptive revisions and 18 of 41 narrative revisions leave no visible line. They say repository channels disagree, neither rate is an involvement rate, and they do not isolate the cause.
Unconfirmed: The authors say they do not isolate why the provenance rates differ and that neither rate is an involvement rate.
- Project A's Co-Authored-By trailer covers 66 of 121 commits that carry any AI trailer.
- Project B: 39 commits revise agent working notes; 1 (2.6 percent) carries the trailer.
- Project A's repository-level rate steps from 2.7 percent to 17.7 percent over time.
- They build an evidence ledger and requirements for recording provenance as it happens.
From the original article
1 min read at huggingface.co
Publishers and funders increasingly ask authors to disclose how AI tools contributed to a piece of research. In computational work the repository is where an author or an auditor would look for the answer, and we ask what it can actually supply.
By Koutian Wu, Zesen Huang, Hejia Geng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.