A new arXiv paper describes an audit of request independence in open Jev-style models for retrieval. Jev's documented interface is designed to evaluate typed questions independently against shared evidence, and the authors set out to test whether that independence actually holds across different checkpoints.
The audit examines seven open Jev-style checkpoints, using 276 constructed three-passage states derived from 92 TREC retrieval queries. This setup allows the researchers to probe how each model handles multiple questions that rely on the same underlying evidence.
The abstract does not report specific findings or conclusions, so the outcome of the audit remains unclear from the announcement alone. The paper's contribution at this stage is the framing and construction of a test for a property that is central to Jev's interface design.