adversarial-verificationlisted

# adversarial-verification adversarial-verification validates claims by making independent agents attack the claim, the evidence, and each other's reasoning. It is a cross-cutting governance engine, not a domain-specific optional helper. Use it when a claim carries real evidence burden: legal, financial, research, grant, patent, investor relations, prior-art, numeric, novelty, consistency, or source-grounded statements. ## Contents - [Core Principles](#core-principles) - [Position In The System](#position-in-the-system) - [Input Shape](#input-shape) - [Output Shape](#output-shape) - [Step 0: Entry Gate](#step-0-entry-gate) - [Step 1: Normalize Claim Card](#step-1-normalize-claim-card) - [Step 2: Assign Agents](#step-2-assign-agents) - [Step 3: Adversarial Rounds](#step-3-adversarial-rounds) - [Step 4: Convergence](#step-4-convergence) - [Step 4.5: Meta-Judge](#step-45-meta-judge) - [Step 5: Return Result](#step-5-return-result) - [External Tool Claim Web Search](#external-tool-claim-web-search) - [Evaluator Artifact Contract](#evaluator-artifact-contract) - [References](#references) - [Gotchas](#gotchas) ## Core Principles 1. Easy convergence is suspicious. For `verification-complexity-level-2` and above, do not check convergence before round 5. 2. There is no mediator during normal rounds. Agents attack, respond, and checkpoint against the claim card and evidence. A meta-judge appears only at deadlock. 3. Concessions must be evidence-bound. The only valid c