Court Closeout Reconciliation

Court closeout reconciliation

You are a deputy/finance clerk producing a closeout package. The inputs always have the same shape and the same trap: local documents carry stale, draft, or mis-labeled values, and your job is to reconcile them against the authoritative court records and the current fee/policy schedules, then write one strict JSON answer. The score is decided by getting the reconciled decisions right and by matching the answer template exactly — not by prose.

Method (do these in order)

1. Read the answer_template first, completely. It is the contract. Note every required_top_level_keys, every enum list, the ordering_rules, and the currency/date/datetime formats. Your output must use enum values verbatim (never substitute prose), numbers to two decimals, ISO dates YYYY-MM-DD and ISO local datetimes YYYY-MM-DDTHH:MM:SS, and each list sorted exactly as the template's ordering rule says.

2. Inventory the inputs. Separate them into (a) local/draft material (finance queue, clerk worksheet/CSV, hearing or closeout notes, intake sheet, petition summary, form excerpts) and (b) the authoritative portal the task points you at (cases, charges, docket-entries, citations, fee-schedules, payment-policies, forms, financial-petitions, and a search facility). Treat the portal as the system of record; treat local material as claims to verify.

3. Pull the portal record for every target matter. For each case/citation/ petition id, retrieve its authoritative row and its associated charges, docket entries, current fee schedule, payment policy, and form metadata for that jurisdiction. Use the search facility when a direct list does not key on the id. The docket disposition line ("recorded with status …", "matter continued; no final order signed") is the ground truth for case status.

4. Reconcile each field with the authority hierarchy (next section), compute any money/date math, and drop anything unsupported.

5. Assemble, sort, and validate against the template before finishing. Verify every enum value is spelled exactly as listed, every list is sorted, item sets are neither over- nor under-populated, and totals equal the sum of the per-item values you actually posted.

See references/field_playbook.md for a field-by-field playbook and references/payment_math.md + scripts/installments.py for the installment math.

Authority hierarchy — who wins a conflict

• Identity (name, DOB), counsel classification, and case status → the portal/CMS. A queued/worksheet DOB, name spelling, or counsel label that disagrees with the portal is corrected to the portal value. A label like APD/PD on a calendar or worksheet does not determine counsel type — use the portal's counsel_type (and the on-record clarification), e.g. "appointed private, county pay" is appointed_private, not public defender.

• What happened in the courtroom (plea, adjudication/conviction, sentence terms, departures, and whether a final order was signed) → the hearing/closeout notes and the docket. These override a stale structured value in the portal charge record. Example: if the bench adjudicated guilt and imposed jail + an assessment, that is a conviction even if a charge row still reads "nolle prosequi"; if the notes say no referral/order was signed, the corresponding datetime/term is empty even if a portal field still carries a tentative value.

• Fee and fine AMOUNTS → the schedule effective on the disposition date. Ignore archived/stale amounts (any schedule row with an end_date in the past), "old worksheet" amounts, and "statutory maximum" notes. Use the current standard/scheduled amount.

When you record a reconciled conflict, pick the resolution_source enum that matches which authority resolved it (portal/CMS, hearing notes, corroborating memo, current fee schedule, or "hold because the order is unsigned").

Fees — post only what is supported

• Post a fee only if the current schedule, a policy, or a signed order supports it. Never add account-management, collection, late-payment, DMV/reinstatement, returned-check, restitution, copy/certification, traffic-school, court-appointed-attorney, or court-reporter charges unless a portal record / current schedule / policy / order directly calls for it. Sticky-note or "do-not-add" reminders are exclusions, not charges.
• A public-defender user fee applies only when counsel is public_defender (not appointed_private, not retained), and only if not waived.
• Mandatory assessments attach to their trigger: a controlled-substance / drug conviction triggers the jurisdiction's current drug-assessment or crime-lab fee; a charge that was amended away to a non-drug offense does not.
• Traffic: add the county surcharge once per citation; use the current standard fine for the tier.
• The register/batch totals are sums over the matters you actually posted; held/excluded matters contribute zero.

Departures, status, and holds

• Record a sentencing departure only when the judge expressly made a departure finding. "Top of the range", "plea agreement cap", or an empty departure reason is not a departure. For misdemeanors, use the template's "not evaluated / not applicable" departure value; for a pending matter, the "not entered / pending" value.
• If a matter has no signed final order (status deferred / continued / pending, or the docket says the order was not signed): do not post financials. Put it in the hold/exclude bucket, use the hold/exclude enum values, set the disposition date to null, and record the next status-check date if the notes give one.

Placeholders — never invent

For a required form field whose value is genuinely absent from all case materials (SSN, driver-license number, address, phone, probation officer/office contact), use the exact placeholder string the materials specify (commonly TBD from case file). Do not guess, and do not borrow a value from a similarly-named party or from search results. List each placeholdered field where the template asks for it, sorted as instructed.

Money math — payment plans / installment orders

Compute, never eyeball. Anchors come from the policy (min/max monthly, first-due offset, down payment, restitution priority, account-fee flag, return-to-court offset) and the balance to be paid:

• total_due = fines_costs_balance + restitution_balance (account fee only if the policy flags it; otherwise excluded and 0).
• full_installments = floor(total_due / monthly); remainder = total_due - full_installments * monthly.
• If remainder > 0: there is one extra final installment equal to remainder, so total_installments = full_installments + 1 and final_payment_amount = remainder. If it divides evenly, total_installments = full_installments and final_payment_amount = monthly.
• first_due_date = submitted/disposition date + policy first-due days (some policies pin it to "the 15th of the next month").
• final_due_date = first_due_date + (total_installments - 1) months (same day).
• return_to_court_date = final_due_date + policy return-to-court offset days.
• Keep the approved monthly within [min_monthly, max_monthly]; classify support by comparing it to disposable income (income - obligations) and the band.
• Payment application order: if restitution > 0 and policy prioritizes it, use "restitution before fines/costs"; if restitution is 0, use "fines/costs only".
• License suspension starts on the conviction date (unless a policy/order names release or petition date); end = start + months.

scripts/installments.py implements this; verify your dates with it.

Output discipline (this is where points are lost)

• Emit exactly the required_top_level_keys, nothing more, nothing less.
• Enum fields: copy the allowed value character-for-character.
• Item sets must match reality: list every matter/charge/exclusion that belongs and omit ones that do not — over-listing an exclusion set or a placeholder set is scored as wrong as omitting one.
• Two-decimal currency numbers; ISO dates/datetimes; null only where allowed.
• Apply every ordering rule (usually sort by case/citation/petition number, and sub-lists alphabetically).
• Cross-check: per-case totals = sum of that case's posted line items; batch totals = sum over posted cases; counts (assessed vs held/excluded) add up.