Reviews

A second, AI-assisted peer review of published papers, with the model, the prompt and the date, checked by a scientist.

Peer review stops at publication. Here a language model reviews published papers a second time, as a demanding referee would: are the claims supported, are the statistics sound, are there missing comparisons, do the numbers in the text match the tables and figures?

Each review states:

the paper reviewed, with a link to the published version;
the model, the date and the exact review prompt or skill used, so anyone can rerun it;
what the model found, separated into points a scientist has checked and confirmed, points checked and rejected, and points not yet checked;
a note of the authors’ reply, when they send one.

Ground rules

Only published papers, preferably open access, and only the parts needed to discuss each point are quoted.
Nothing is published until a scientist in the field has checked every point marked as confirmed. Unchecked model output is never presented as a finding.
Authors are informed and can reply. Their response is added to the review.
The goal is to test what AI review can and cannot do, not to grade people.

First reviews in preparation

The review prompt and format are being defined. Suggest a paper.