Reproducibility checks during machine-learning peer review often stumble because reviewers and authors do not share the same expectations about what the code is supposed to reproduce. The paper argues that this shared understanding is a prerequisite for meaningful checks.

To address this, the authors present an author-in-the-loop collaborative audit. Instead of treating reproducibility as a purely automated pass/fail test, the method makes authors' declarations explicit inputs to an AI system. This allows the audit to be guided by what the authors themselves say the code should reproduce.

The abstract describes the approach as agentic and collaborative, but it does not detail the exact mechanics or results. The work appears to target a machine learning conference, though the specific venue is not named in the abstract.