Physics Reviewer

An AI-assisted peer review system for scientific papers in condensed matter physics. The agent verifies whether a paper is trustworthy through a structured 13-stage pipeline — from PDF reading to final review report.

Quick Start

Prerequisites

• A paper PDF to review
• An LLM that supports skill invocation (Claude Code, Codex, Copilot CLI, etc.)

First Review

# 1. Create a workspace for your paper
cp -r workspace/paper_template workspace/paper_001

# 2. Place the paper PDF
cp your-paper.pdf workspace/paper_001/manuscript/paper.pdf

# 3. Fill in metadata
# Edit workspace/paper_001/metadata/metadata.md — set Journal and Round

# 4. Invoke the skill
# Tell your LLM: "Review this paper for PRB, first round"
# The agent loads skills/first-review.md and runs the pipeline

Revision Review

Place the response letter and revised manuscript in workspace/paper_001/revision/, then invoke the revision review skill.

How It Works

The first review pipeline has 13 stages organized in three phases:

Phase A — Foundation (sequential):

• Stage 1: Convert PDF to Markdown, read thoroughly, write structured summary
• Stage 2: Extract every scientific claim, classify by type and priority
• Stage 9: Literature comparison — web search for prior work to verify novelty

Phase B — Verification (parallel):

• Stage 3: Proofreading and language check
• Stage 4: Formula verification — variable definitions, dimensional analysis, boundary conditions
• Stage 5: Mathematical reproduction — step through every derivation
• Stage 6: Numerical verification — recompute from stated inputs
• Stage 7: Logic verification — extract reasoning chains, detect gaps and overclaiming
• Stage 8: Physics verification — symmetries, conservation laws, limiting cases
• Stage 10: Experimental verification (N/A for pure theory)
• Stage 11: Code review and reproducibility assessment
• Stage 12: Figure verification — labels, captions, text-figure consistency
• Stage 14: Completeness analysis — what is missing from the paper as a whole
• Stage 15: Deep physics probing — web-driven search for domain controversies, open questions, and alternative mechanisms

Phase C — Synthesis:

• Stage 13: Risk analysis — aggregate all findings, classify severity, score rubrics, generate the review report
• Stage 13b: Quality check — independent subagent validates the review against benchmarks

All analysis outputs go to analysis/. The final review goes to report/review.md.

What the Review Covers

• Mathematical correctness — every derivation, every formula, every variable definition
• Numerical accuracy — recompute from stated inputs, check significant figures
• Physical validity — symmetries, conservation laws, limiting cases, DFT parameter audits
• Logical coherence — does each conclusion follow from its premises? Are there hidden assumptions?
• Literature novelty — web search for prior work, citation accuracy, missing references
• Completeness — what should be in the paper but isn't? Missing examples? Missing data?
• Presentation — spelling, grammar, notation consistency, figure quality

Directory Structure

physics-reviewer/
├── skills/              # Review workflow skills
│   ├── first-review.md       # 13-stage first review pipeline
│   ├── revision-review.md    # R0-R5 revision review pipeline
│   └── adversarial-review.md # A1-A7 adversarial review pipeline
├── knowledge/           # Reference knowledge (journal profiles, physics)
│   ├── journals/             # Nature, Science, PRL, PRX, PRB, etc.
│   └── physics/              # DFT, magnetism, symmetry, transport
├── rubrics/             # 1-5 scoring rubrics for 6 dimensions
├── templates/           # Output templates for review reports
├── reviewer_memory/     # Reviewer profile configuration (customize to your style)
├── examples/            # Example reviews for quality benchmark calibration
└── workspace/           # Paper workspaces (one per paper)
    └── paper_template/  # Template to copy for new reviews

Customizing

Edit files in reviewer_memory/ to match your reviewing style:

• profile.md — your review philosophy, decision thresholds, report structure
• language_patterns.md — your phrase templates for openings, criticisms, suggestions
• focus_areas.md — your checklist of dimensions to evaluate

Journal Support

Journal profiles in knowledge/journals/ provide journal-specific thresholds and rejection triggers. Currently supported: Nature, Science, PRL, PRX, PRB, APL, Advanced Materials. Add new profiles by following the existing format.

License

MIT