metadata:
  title: AI agent evaluation release gate
  version: 1.0.0
  license: Apache-2.0
  maintainer: Signal Studio
  updated_at: 2026-08-26
  guide_url: https://signalstudio.cc/guides/turn-agent-evaluation-results-into-a-production-quality-gate/
  book_url: https://signalstudio.cc/books/evaluating-ai-agents/
  boundary: This example structures a release decision. It does not supply universal thresholds or prove production safety.

release_candidate:
  id: SYNTHETIC-agent-2026-08
  model_version: synthetic-model-v2
  workflow_version: refund-review-2.1.0
  task_portfolio_version: refund-portfolio-1.4.0
  environment_version: refund-sandbox-1.2.0

offline_evidence:
  repeated_trials_per_task: 20
  accepted_outcome_rate: 0.94
  serious_failure_count: 0
  unresolved_high_risk_cases: 2
  reviewer: domain-expert-required

decision:
  status: hold
  reason_codes:
    - unresolved-high-risk-cases
  decision_owner: release-owner
  approved_at: null

canary:
  eligible_traffic_percent: 1
  allowed_actions: [read, propose]
  prohibited_actions: [execute-financial-effect]
  duration_hours: 24

live_monitors:
  - serious_failure_rate
  - unknown_effect_state_rate
  - human_escalation_rate
  - cost_per_accepted_outcome

stop_and_rollback:
  stop_on_any_serious_failure: true
  stop_on_unknown_effect_state: true
  rollback_owner: platform-on-call
  recovery_runbook: governed-internal-reference