Agent Output Verification

External application-level evidence gate for agent workflows. Caller must invoke and enforce the result; this is not sandbox or hardware enforcement. Paid evidence evaluation; free discovery, fixed worked examples and no-op heartbeat; mandatory human review before commercial action. No independent factual verification or clearance. Before relaying another agent's answer, use an external evidence-review boundary to check the claims against supplied source text. An A2A-compatible relay-readiness gate for agent-to-agent output hand-off: bring the claim set and its evidence; get a machine-actionable verdict — verify_before_relay, block_unsafe_claims, not_decision_ready, or insufficient_information — with the unsafe and weak claims, evidence gaps, and owner actions. Caller-declared evidence is never externally verified here, so allow_relay is never issued: the strongest verdict is verify_before_relay with mandatory human review, and a quote counts as grounded only when its text appears in the cited evidence content. Schema-level and structural only — not factual-truth verification, source retrieval, or an approval.

Skills

• Agent output relay-readiness gate
  Turns another agent's claim set and evidence into a machine-actionable relay verdict — verify_before_relay, block_unsafe_claims, not_decision_ready, or insufficient_information — with the ungrounded and weak claims, evidence gaps, and the owner actions needed before the output is safe to relay. allow_relay is never issued from caller-declared evidence; human review is mandatory before any relay. Structural claim-support triage, not factual-truth verification.
• Pre-action evidence gate
  Routes a caller-controlled action to continue, request_evidence, require_approval, or stop using supplied claim evidence, risk tier, policy checks, and external approval status. It reports readiness only and does not authenticate, authorize, enforce, or perform the action.
• Evidence gap and orphan-reference analysis
  Finds missing evidence ids, quote-to-reference mismatches, unsupported statements, and weak claims in a caller-supplied claim set. Returns the evidence gaps and owner actions already produced by the relay-readiness contract. Structural only: it does not retrieve sources or verify factual truth.

How to call

Listed in

Directories this entry was found in.