# Adversarial review

A reusable skill the "Clear security questionnaires without guessing" recipe pulls in. Add it once and every agent on the team can use it.

Refute every drafted answer before it ships, so nothing overstates the company's real security posture.

## Method
1. Spawn a second-pass sub-agent that did not write the original drafts.
2. Give it only the answer library and the policy docs — not the drafter's reasoning.
3. Have it independently re-derive an answer for each question from the library and policy docs alone.
4. Compare its re-derived answer against the drafted answer and flag any mismatch.
5. Reject any drafted answer that overstates a control, invents a certification, or lacks a citation the reviewer can verify.
6. Return each answer as approved, revise, or open — never a silent pass-through.

## Rules
- The reviewer works from source documents only, never from the drafter's confidence or phrasing.
- Reject on overstatement even when the underlying fact is directionally true.
- Reject on missing evidence — a plausible answer without a verifiable citation does not pass.
- On disagreement between reviewer and drafter, escalate to the review queue instead of guessing.

## Output
A per-question verdict list: approved, revise, or open, each with the reviewer's stated reason.