Adversarial Doc Review

Quick start

1. Get the document (path/URL/pasted text) plus the intended audience and stakes.
2. Build a claim ledger: every definition, assumption, factual claim, normative claim, control claim, and route-authorization claim must be traceable to a location in the document.
3. Stress-test internally for consistency, logic, edge cases, evidence, citations, non-vacuity, and specification gaming.
4. Verify externally by default: use a web browsing/search tool to fact-check time-sensitive and externally-checkable claims.
5. When the document authorizes research evidence or stage advancement, apply references/epistemic-assurance-contract.md and test whether labels correspond to actual properties.
6. Produce a structured report in chat by default using references/report-template.md and references/review-checklist.md. If a file is requested, write <docname>.review.md next to the document.

Workflow decision points

• If web browsing is unavailable or disallowed, ask for permission or user-provided sources; otherwise label items as unverified and state what would verify them.
• If the document is long, ask which sections matter; still scan the rest for contradictions, definitional drift, unsupported strong claims, and stage-advancing statements.
• If the document claims validation, verification, independence, immutability, isolation, replay, auditability, or route readiness, inspect the mechanism and evidence rather than accepting the label.
• Default deliverable: structured report in chat. If a file is requested, write <docname>.review.md in the same folder as the input document.

Inputs supported