adversarial-review
Adversarial Review
Overview
Score-based reviews list weaknesses by severity and rarely commit. This skill runs the opposite pass. It forces one reviewer to write the single argument an area chair would reject the paper on. A fresh adjudicator then rules that argument point by point. The verdict is computed from the rulings by a helper, so the adjudicator never grades its own work.
The exercise surfaces one specific failure mode that balanced reviews miss: the headline-killing objection. A balanced reviewer files scope-overclaim as one major item among five and moves on. An adversarial reviewer must pick the most damaging line and develop it in about 200 words. That forced commitment produces sharper feedback than a ranked list.
Core principle: commit to one argument, then judge it externally. The attack must not hedge. The verdict must not be self-graded.
This is detect-only. The skill reads .writing/manuscript/*.tex and writes one memo plus a progress row. It never edits manuscript prose and never auto-mutates .writing/ claim state. The verdict is advisory: surface it to the user and let the user decide what to fix.
Relation to the two-stage review gate. Drafting already runs spec review and manuscript quality review per section (see skills/planning-foundation/references/review-loop-protocol.md). Those gates check on-spec coverage and prose quality. This skill is different. It runs once on a whole stable draft and asks the headline-rejection question that per-section review cannot see.