review

# /review — Manuscript Review ## Modes - **external** (default): review another author's work as submitted. - **own-work**: review the user's own draft. The clean room is real, not declared: spawn ONE fresh-context subagent whose prompt contains only the manuscript text (and explicitly listed evidence files), and relay its findings. Never present a same-session read-through as a clean-room review — the drafting conversation is context contamination by definition. ## Evidence boundary Judge the manuscript **as submitted**. Do not use prior chat memory, unstated author intentions, or model background knowledge as evidence for or against a claim. If a claim cannot be assessed from the manuscript and its cited sources, say so — that is itself a finding. ## Findings format Every finding is anchored (section/paragraph or quoted span) and typed: | # | Type | Anchor | Finding | Severity | |---|------|--------|---------|----------| Types: `claim-exceeds-evidence` | `gap-contribution-mismatch` | `method` | `structure` | `clarity` | `citation`. Split findings three ways: defects that block the recommendation, improvements that would strengthen it, and questions the author must answer. Consult `references/argument-checklist.md` for the interrogation checklist and examiner-attack pre-mortem when the review targets argument quality. ## Recommendation vocabulary Exactly one of: `accept` | `minor_revision` | `major_revision` | `reject_resubmit` | `reject` | `no_recommendati...