Adversarial verification for AI agents: an independent skeptic attempts to refute a claim before the agent acts on it, returning a structured verdict (refuted / supported / inconclusive) with calibrated confidence, findings, citations, and an Ed25519-signed receipt on paid tiers.