{
  "world_state_before": "Calvin currently processes the same work inbox manually, with a seven-day baseline of 42 minutes per workday and no delegated triage rules.",
  "intervention": "Allow an email agent to classify and draft responses for seven workdays while Calvin reviews every action before anything is sent or archived.",
  "unknown_outcome": "It is unknown whether sorting time saved by the agent will exceed setup, review, repair and error-recovery time without hiding an important message.",
  "evidence_required": [
    "manual baseline ledger",
    "daily agent runtime and human review ledger",
    "intervention and repair log",
    "missed-message audit",
    "setup-time receipt"
  ],
  "failure_story": "If the net saving is zero or negative, or an important message is mishandled, that result remains the episode and determines the next safety test.",
  "world_state_after": "The inbox world gains measured review costs, known failure cases, revised triage rules and a specific unresolved question for the next run.",
  "unresolved_assumptions": [
    "The seven days are sufficiently comparable to the manual baseline.",
    "The missed-message audit can detect material errors."
  ],
  "decision": "ALLOW_PLANNING",
  "release_authority": "A3"
}
