research-reviewer

Research Reviewer — Iterative Blind Peer Review

Run an iterative, multi-agent review loop that drives a manuscript to a flawless 100/100. Each iteration spawns a fresh Reviewer Agent with no memory of prior drafts so the review stays unbiased and never "overfits" to earlier versions.

How to run

1. Locate the manuscript. Ask the user for the paper (a .tex/.md/.pdf path, a folder, or pasted text) plus its figures/tables/datasets if available. Do not proceed without the actual content to review.

2. Spawn a fresh Reviewer Agent (blind). For each iteration, launch a new subagent via the Agent tool (general-purpose). Pass it:

   • the full system instructions from reviewer-agent.md (read it, inline it),
   • the current manuscript content only — never the previous review, score, or change log. This is what keeps the "fresh pair of eyes."

3. Collect the review. The agent returns: a score X/100, itemized deductions mapped to the 4 pillars, refactored/polished text for the flawed sections, and a Required Experiments checklist.

4. Apply fixes. Edit the manuscript with the agent's refactored sections and notation fixes. Surface the Required Experiments checklist to the user — those need real runs and cannot be auto-resolved.

5. Loop. Re-spawn a brand-new Reviewer Agent on the updated manuscript. Repeat until the score is 100/100 with zero remaining deductions, or the only blockers left are Required Experiments the user must run. Report the score trajectory each round.

The four pillars the reviewer enforces

1. Methodological Soundness & Mathematical Precision — assumptions stated, notation consistent, every design choice justified.
2. Empirical Rigor & Validity of Claims — error bars / CIs / p-values, fair baselines, clean single-variable ablations.
3. Scope of Validity & Narrative Honesty — no overclaiming, explicit limitations.
4. Visual Layout & Editorial Integrity — 30-second graphic test, no collisions, valid citations.

Notes

• The full reviewer persona, scoring protocol, and output format live in reviewer-agent.md (same folder). Read it and pass it verbatim to each subagent.
• Keep iterations genuinely blind: a new Agent call starts fresh — do not reuse a prior agent via SendMessage, and do not leak prior scores into the prompt.
• If the paper has named benchmarks (e.g. method must beat specific baselines), tell the reviewer to enforce them — see the "Paper-specific benchmarks" section in reviewer-agent.md.