docs/spec/tasks/JE-08.json

{
  "id": "JE-08",
  "title": "Accrual recomputation and reversal integrity",
  "tier": "L2",
  "family": "Accrual mechanics",
  "sources": [
    "DS-A",
    "DS-J"
  ],
  "mapping": "Generate month-end accrual entries from CT invoice/payment date gaps (goods received before period end, paid after). Support pack carries the accrual calculation schedule. Seed reversal failures in the following period.",
  "scenario": "Period-end accruals were booked and should have reversed in the following period. Some did not.",
  "brief": "Verify each accrual is correctly calculated from its schedule and that it reversed completely and in the correct period. Flag every failure.",
  "tools": [
    "query_ledger",
    "get_entry",
    "list_documents",
    "open_document",
    "recompute",
    "compare_period",
    "disposition",
    "submit"
  ],
  "groundTruth": "Seeded across 25 accruals: 4 never reversed, 3 reversed in the wrong period, 2 reversed at a different amount than accrued, 3 whose booked amount does not agree to the attached calculation schedule, 2 double-counted (accrued and also invoiced in the same period).",
  "verifier": [
    "flagged set and issue_type matched against seed",
    "for schedule-mismatch cases, the agent's stated correct amount must equal the schedule total within $1",
    "compare_period or an equivalent cross-period query must appear in the trajectory before any reversal-related flag",
    "the double-count cases require BOTH the accrual and the invoice entry cited in evidence_ids"
  ],
  "judgeFocus": [
    "Did the agent recompute from the schedule, or accept the booked amount as given?",
    "For the double-count, is the rationale about duplication of expense, not merely 'unusual timing'?"
  ],
  "dials": [
    "accrual population size",
    "schedule complexity (rates, proration, headcount drivers)",
    "partial reversals",
    "reversal booked to a different account"
  ],
  "guard": "Schedule totals must require at least one arithmetic step — never print the booked amount on the schedule.",
  "weight": {
    "outcome": 0.5,
    "process": 0.2,
    "rationale": 0.3
  },
  "fixtures": {
    "real": [
      "DS-A: invoice_date vs payment_date gaps across period boundaries; multi-year history FY2020–FY2027 for prior-period comparison"
    ],
    "synth": [
      "Accrual entries + reversal entries",
      "Accrual calculation schedules (multi-step, total not pre-printed)",
      "Seeded reversal failures"
    ]
  }
}