Distilling a Nature Remote-Sensing Reviewer

Last week a few of us met for dinner. Before all the dishes arrived, the conversation had already reached the familiar subject: Reviewer #2.

Peer review has a peculiar ability to produce surprises late in the process. A reviewer may question the reliability of the data, decide that the method is not sophisticated enough, ask for stronger validation, or request a new block of experiments in the second or third round. Opening a decision letter can feel uncomfortably close to opening a blind box.

That led to a simple idea: could the reviewer be distilled?

More precisely, could we build a cyber-reviewer from the public review record of high-level journals? Rather than waiting until after submission to discover the obvious vulnerabilities in a manuscript, an AI-assisted reviewer could be used as a pre-submission stress test.

I began with remote sensing, the area I know best. I collected publicly available Peer Review Files from recent remote-sensing-related papers in Nature and Nature Communications, extracted the reviewer comments, cleaned and segmented them, and then organized recurrent patterns of criticism. The result is a Nature-style remote-sensing reviewer skill: nature-remote-sensing-reviewer-skill.

I have tested it on my own material. It is not a substitute for a real reviewer, but the result is recognizably close to the kinds of questions experienced reviewers ask—occasionally including the familiar art of the politely phrased second blow.

The skill is available on GitHub: nature-remote-sensing-reviewer-skill. Remote sensing was only the first test case; the same corpus-distillation approach can be extended to other research areas.