Erp Finance Close Reviewer

ERP Finance Close Reviewer

This skill turns the shared ERP finance API into correct, template-shaped JSON answers. The rules below were reverse-engineered from real task inputs and their official answers and re-verified against the live API. Follow them literally; the grading is exact-match on values, enums, rounding, and sort order.

1. API access (always do this first)

• Base URL: http://127.0.0.1:8029. Ignore any other base URL a prompt mentions (e.g. :8005, :8029 variants in narrative text). Confirm with GET /health.
• Fetch only via HTTP GET + curl. Never read local server files.
• List endpoints wrap rows: {"endpoint","count","total","offset","limit","data":[...]}. Read data. Default limit=100, max 500. If total > count, paginate with offset, or just pass limit=500. Always check total to be sure you got everything.
• Exact-match filters by field name work on list endpoints, e.g. /api/ap/bills?claim_id=CLM-2025-0001, /gl/balances?account=1250&period=2025-03.
• Single-object endpoints exist for claims (/api/claims/{id}) and compliance (/api/compliance/{profile|ownership|registry|screening|bank|risk}/{business_id}).

Endpoint map

Pitfall — bill_id is NOT unique. The same bill_id can return multiple rows with different vendor/amount/claim_id. Always disambiguate by also matching claim_id, vendor_id, and amount. The same applies in /api/ap/aging.

2. Reimbursement / AP close & stale-snapshot review (claim batches)

Task types: "reimbursement-to-AP close" (train_001 family) and "stale AP snapshot reconciliation" (train_004 family). Both take a list of candidate claim_ids and ask which can stay in the AP batch.

Per-claim procedure

For each candidate claim:

1. GET /api/claims/{claim_id} → record status, amount, vendor_id.
2. GET /api/ap/bills?claim_id={claim_id} → its linked bill(s).
3. Match a valid bill: a bill is the claim's valid bill only if bill.amount == claim.amount AND bill.vendor_id == claim.vendor_id AND bill.status != "void". (Claim vendor_id may be null; then a bill with a real vendor does NOT match → mismatch.)
4. For the matched bill, GET /api/ap/payments?bill_id={bill_id} and keep only payments whose bill_id+vendor_id line up with this claim's bill.

Disposition decision tree (apply top-down)

• claim.status not in {approved,paid} → blocked (reason block_unapproved_claim).
• No bill at all for the claim → blocked (no AP evidence).
• Matched bill is void → blocked (reason ignore_void_bill).
• Bill amount ≠ claim amount OR bill vendor ≠ claim vendor (and not void) → blocked (reason exclude_amount_or_vendor_mismatch).
• Matched valid bill with status=="paid" AND a cleared payment for the full claim amount → paid (reason replace_with_matched_paid_bill when a stale snapshot listed a different/scheduled bill).
• Matched valid open bill (scheduled/approved), payment exists but is processing/ scheduled (not cleared) → payable, in-flight (reason mark_in_flight_payment).
• Matched valid open bill, no payment, snapshot agrees → payable (reason current_snapshot_ok).

"eligible to stay in batch" = payable ∪ paid. "blocked"/"not_ready" = everything else. In train_001 the template separates paid from payable; in train_004 both are eligible. Read the template to see which split it wants.

AP open balance (CRITICAL: cleared-payments-only)

Compute the open balance yourself — do NOT trust /api/ap/aging balance/paid_amount, because aging counts processing/scheduled payments toward paid_amount, hiding real open balances.

open_balance(claim) =
  0.00                                   if no matched valid bill, or bill status == "paid"
  bill.amount − sum(payments where status == "cleared")   otherwise

Round to 2 decimals. Example: an in-flight reimbursement (bill scheduled, payment processing) has open balance = full bill amount even though aging shows balance 0.

• ap_open_balance_total (train_001) = sum of open_balance over payable claims only.
• ap_balance_by_claim (train_004) = map every candidate claim → its open_balance.

Stale-snapshot correction enum

Map each candidate claim to its reason from the disposition tree above. Allowed values: current_snapshot_ok, mark_in_flight_payment, replace_with_matched_paid_bill, exclude_amount_or_vendor_mismatch, ignore_void_bill, block_unapproved_claim.

Close-log requirement (train_004 family)

When an in-flight payment / replaced-paid-bill correction means AP must be re-posted, find the supporting close log: the AP-area log (area=="AP") for the period of that bill (e.g. bill_date 2025-04-* → period 2025-04) documenting the journal entry (message ~ "Manual journal entry posted"). Return close_log_required.required=true and its log_id(s) sorted ascending. If no correction needs posting, required=false, ids=[].

batch_status enum

• train_001 (ready_to_close / open_payables / blocked): blocked if any item blocked; else open_payables if any valid unpaid AP bill remains; else ready_to_close.
• train_004 (ready_to_send / needs_ap_refresh / blocked): blocked if a candidate is hard-blocked in a way that stops the batch; else needs_ap_refresh if any candidate required a stale-snapshot correction (in-flight/replaced/void/mismatch) but eligible items remain; else ready_to_send.

3. Compliance gates: onboarding release & AP payment release

Two task types share the compliance endpoints, keyed by business_id: fetch profile, ownership, registry, screening, bank, risk. Find the linked vendor via profile.vendor_id then /api/vendors?vendor_id=... for vendor status.

Shared field definitions

• bank mismatch: bank.bank_account_status == "name_mismatch". (closed is separate.)
• invalid tax id: profile.tax_id (or registry.tax_id) is NOT ^TIN\d{6}$, OR equals the placeholder TIN999999. (Letters like TIN12X899 → invalid; all-9s → invalid.)
• expired license: compare registry.license_expiry to the task's as_of date. Default rule (payment-release / train_005): license_expiry < as_of_date (plain). Read the answer_template — it states the comparison date. (Onboarding/train_002 behaved like a ~30-day grace: license_expiry < as_of − 30d; when in doubt use the template's comparison_date.) If license is in profile.missing_fields, treat it as a missing document, not an expired license.
• screening not run: screening.sanctions_check_status == "not_run" (also surfaces as the screening_not_run hard stop).
• confirmed PEP: screening.pep_status == "confirmed_pep" (possible_pep does NOT count).
• sanctions confirmed: screening.sanctions_check_status is a confirmed hit.
• shell company: ownership.shell_company_suspected == true.
• vendor on hold: linked vendor status == "on_hold".
• missing required documents: profile.missing_fields non-empty.
• risk score override: risk.risk_score >= 70.

3a. Vendor onboarding intake (train_002 family)

Output: per-business decision, reportable_ubo_counts, hard_stop_flags, follow_up_business_ids, overall_release_ready.

reportable_ubo_counts: from ownership.ubo_list, count unique owner names with ownership_pct >= 25. Dedupe by name (the same name listed twice counts once); owners below 25% are excluded.

hard_stop_flags (per business, alphabetical, empty list if none). Allowed enum: bank_closed (bank_account_status=="closed"), bank_name_mismatch, confirmed_pep, expired_license (only when license present & expired, see grace note), sanctions_confirmed, screening_not_run, shell_company_suspected, missing_required_documents, vendor_on_hold.

decision per business:

• approve — no hard-stop flags.
• awaiting_information — only curable flags present: missing_required_documents and/or screening_not_run, and nothing more serious.
• escalate — any serious flag present (confirmed_pep, sanctions_confirmed, bank_name_mismatch, bank_closed, expired_license, shell_company_suspected, vendor_on_hold). Serious-flag presence dominates even if curable flags also exist.

follow_up_business_ids: every business with ≥1 hard-stop flag (i.e. decision ≠ approve), ascending. overall_release_ready: true only if every business is approve.

3b. AP payment release after account changes (train_005 family)

Output: decisions, bank_mismatch_ids, invalid_tax_ids, expired_license_ids, review_queue_ids, risk_score_override_flags. Echo task_id/batch_id/as_of_date/ target_business_ids exactly as the template/payload requires.

decision per business (enum release/hold/escalate):

• escalate — identity-integrity failure: invalid_tax_id OR confirmed_pep OR sanctions hit.
• release — no issues at all (bank verified, valid tax, license current, screening run, pep none, sanctions clear). Note: risk.review_status like in_review and a high risk_score alone do NOT block a release if everything else is clean.
• hold — has operational issues (bank name_mismatch/closed, expired license, screening not run, missing docs, risk_score ≥ 70) but identity is intact (valid tax, no confirmed pep, no sanctions hit).

ID lists (each ascending by business_id):

• bank_mismatch_ids = bank_account_status == "name_mismatch".
• invalid_tax_ids = invalid tax id rule above.
• expired_license_ids = license_expiry < as_of_date (this family uses plain comparison; template confirms comparison_date: as_of_date).
• risk_score_override_flags = risk_score >= 70.
• review_queue_ids = every business that is NOT released (i.e. decision ∈ {hold, escalate}).

4. Prepaid-to-GL close reconciliation (train_003 family)

Inputs: an entity, a close period YYYY-MM, a scoped list of prepaid invoice IDs, the accounts to reconcile, and a variance_threshold_abs. Use straight-line amortization as encoded in each invoice (monthly_amortization, original_amount, service_start/end).

Per-invoice math (round each money value to 2 decimals)

Let months_elapsed = (period_year − start_year)*12 + (period_month − start_month) + 1 (inclusive of both the service-start month and the close-period month), capped at the invoice's total service months (end_y−start_y)*12 + (end_m−start_m) + 1, and floored at 0.

march_amortization        = monthly_amortization          (the period's monthly figure)
cumulative_amortization   = min(monthly_amortization * months_elapsed, original_amount)
ending_balance            = round(original_amount − cumulative_amortization, 2)

Keep tiny rounding residuals as-is (e.g. ending 0.01); do NOT force to 0. A fully-elapsed invoice (months_elapsed ≥ total months) ends at original_amount − original_amount.

Per-invoice flags

• default_missing_term_flag = "missing_contract_dates" in data_quality_flags.
• exception_flag = data_quality_flags is non-empty (ANY flag, e.g. rounded_amount OR missing_contract_dates). Emit invoice_results in the same order as the scope list (not sorted).

Account rollup (per account key, e.g. "1250", "1251")

• account_name from the GL balance row. selected_invoice_count = scoped invoices on that account.
• original_amount_total, march_amortization_total, cumulative_amortization_through_march, schedule_ending_balance = sums of the per-invoice values for that account (2 decimals).
• gl_ending_balance = GET /api/prepaids/gl-balances?account={acct}&period={period} → ending_balance.
• variance_amount = schedule_ending_balance − gl_ending_balance (round 2).
• variance_flag = abs(variance_amount) > variance_threshold_abs.
• has_default_missing_term_flag = any invoice on the account has default_missing_term_flag.
• account_status enum: requires_reconciliation when variance_flag is true (material variance); else variance_review when no material variance but data-quality/missing-term exceptions exist; else reconciled.

Top-level lists

• selected_invoice_ids = scope list, same order.
• default_missing_term_invoice_ids and exception_invoice_ids = the matching invoice IDs, ascending.

5. Output discipline (applies to every task type)

• Return only the JSON object that conforms to answer_template.json. No prose.
• Honor required_top_level_keys / top_level_order and additional_properties_allowed.
• Currency: USD, 2 decimals (cents). When a template says "USD cents", confirm whether it wants integer cents vs a 2-decimal number — match the template's unit/type.
• Sorting: every ID list ascending by ID unless the template says otherwise (prepaid selected_invoice_ids/invoice_results keep scope order).
• Enums: use the exact allowed strings; never invent values.
• Re-read the template's per-field description/definition before finalizing — it often states the exact computation (e.g. "schedule_ending_balance minus gl_ending_balance", "risk_score >= 70", "bank_account_status is name_mismatch").

6. Quick SOP per task type

1. Identify task type from the prompt + payload (claim batch → §2; business batch with onboarding/intake → §3a; business batch with account-change/payment-release → §3b; prepaid invoices + accounts + period → §4).
2. Read the payload (scope/batch IDs, as_of/period, thresholds) and the answer_template.json (keys, enums, ordering, precision).
3. Pull every needed record from the API at http://127.0.0.1:8029 (paginate to total).
4. Apply the rules above; compute money to 2 decimals; dedupe and sort as specified.
5. Emit JSON exactly matching the template; do a final pass against each field's description.