{
  "status": "Illustrative acceptance plan, not an executable test suite or observed production result",
  "topic": "AI action rollback and compensation",
  "date": "2026-11-04",
  "sourceArticle": "https://aptixlabs.com/insights/ai-agent-rollback-design",
  "setup": [
    "Use synthetic accounts and records",
    "Record the environment and implementation version",
    "Agree the expected product contract before testing"
  ],
  "reviewSequence": [
    "Mark the commit point of every side effect",
    "Define compensation and its accountable owner",
    "Test partial completion across dependency boundaries"
  ],
  "cases": [
    {
      "id": "1",
      "condition": "Database write succeeds but notification fails",
      "expectedBehavior": "Expose the chosen recovery state",
      "observedBehavior": null,
      "status": "not_run"
    },
    {
      "id": "2",
      "condition": "Compensation is denied",
      "expectedBehavior": "Escalate without hiding the original action",
      "observedBehavior": null,
      "status": "not_run"
    },
    {
      "id": "3",
      "condition": "User repeats a recovery request",
      "expectedBehavior": "Do not compensate twice",
      "observedBehavior": null,
      "status": "not_run"
    }
  ],
  "tradeoff": "Compensation can itself confuse users if the original action was visible. Sometimes the best recovery is an explicit correction and human follow-up rather than an automatic chain of reversals."
}