As Gemini moves into scientific workflows, how do we preserve evidence of what the model actually contributed?
We audited 569 Git commits (466 and 103) from two research repositories assisted by Claude Code and Cursor. Gemini was not one of the audited tools; the point for this forum is that the same failure will apply to Gemini-assisted work, because it is a failure of Git, not of any one model. In one repository the Co-Authored-By trailer covered only 66 of 121 commits carrying any AI trailer. In the other, 39 commits revised a file holding an agent’s working notes and only 1 (2.6%) carried the trailer.
The signals were unreliable. Trailers appear inconsistently, absence proves nothing, and simple counts upgrade easily into false authorship claims.
We built a small open tool, Git Provenance Auditor, that produces a Provenance Evidence Ledger stating only what the history supports, and explicitly listing what it does not support.
Paper: [2610.28655] What Git Does Not Record: Provenance Legibility in Two AI-Assisted Research Repositories (the evidence ledger is included as ancillary files on arXiv)
Koutian Wu,