reflect-3-attempt-01

name	reflect-3_attempt_01
description	SOP for ProcureOps ERP procurement/receiving/AP tasks — pull records from the read-only API as source of truth, apply sourcing/receiving/three-way-match/AP-release rules, and emit answer.json that exactly matches the task template.

ProcureOps Procurement / Receiving / AP Solver Skill

You answer ERP "decision packet" tasks in the ProcureOps domain (sourcing nomination, receiving closeout, AP close/hold-release, change-control, chargeback netting). Each task gives you input/prompt.txt, local payload files (memos, registers, watch sets), and an answer_template.json. Your job: gather the authoritative records, apply the business rules, and return ONE JSON object that matches the template exactly.

---

1. The remote API is the source of truth

Base URL (ignore any localhost/127.0.0.1:8006 in the prompt — use this):

<remote-env-url>

GET-only. Collections: programs, suppliers, items, contracts, purchase_requisitions, purchase_orders, receipts, ap_invoices, payments, approval_events, budget_snapshots, vendor_risk_events.

Patterns:

• GET /<collection> -> {"count": N, "results": [...]}
• GET /<collection>/<id> -> single object (404 if missing)
• Filter by any record field (case-insensitive), e.g. GET /receipts?po_id=PO-AX17-4481, GET /ap_invoices?po_id=..., GET /payments?invoice_id=..., GET /vendor_risk_events?supplier_id=..., GET /purchase_orders?contract_id=..., GET /contracts?sku=....
• start/end filter the collection's primary date field.

Rules of engagement:

• Local payload files name the scope (target IDs, watch set, business request, chargeback register, tax rate). The API gives the values. When they disagree, trust the API for record values; trust the local register for things that ONLY exist locally (e.g., chargeback approvals, tax rate, opening balances).
• A filtered query often returns MORE than the one record you expect (e.g. a PO can have several receipts and several invoices, some old/paid/cancelled). Always fetch the full set and then apply the task's date/scope filters yourself. Do not assume one-to-one.
• Verify each target ID actually resolves; if a referenced ID is absent, treat it as "missing receipt / no record" (that is itself a decision input), and use the fallback IDs the payload provides.

Record shapes (key fields)

• program: budget_cap, committed_amount, owner, status, priority.
• supplier: risk_rating (low|medium|watch|...), status, name, payment_terms.
• contract: status, price_type, unit_price, ceiling_amount, sku, supplier_id, program_id, effective/expiry_date.
• purchase_order: status (open|confirmed|partial_receipt|received|closed| cancelled), lines[] (line_id, sku, quantity, unit_price), subtotal, tax, total, contract_id (may be null), requisition_id.
• receipt: status (accepted|accepted_with_note|inspection_hold|...), lines[] (po_line_id, sku, quantity_received, quantity_rejected, inspection_status), receipt_date, packing_slip, receiver, warehouse_id.
• ap_invoice: status (approved|on_hold|pending_receipt|paid|...), hold_code (QTY_VARIANCE|PRICE_VARIANCE|NO_RECEIPT|null), lines[] (po_line_id, sku, quantity_billed, unit_price), subtotal, freight, tax, total, receipt_id (may be null).
• payment: invoice_id, amount, scheduled_date, status (scheduled|...).
• approval_event: object_id, object_type, action (submitted|approved|...), actor, event_date, event_id, note_code.
• budget_snapshot: snapshot_id, budget_cap, committed_amount, pending_invoice_amount, snapshot_date.
• vendor_risk_event: , (low|medium|high|critical), (open|monitoring|closed), , , , .

---

2. Output conventions (these decide most of your score)

1. Match the template structure EXACTLY. Emit every required key, with the exact key names, nesting, and enum spelling shown in the template. Return ONLY the JSON object — no prose.
2. Copy literal record values verbatim. Fields like invoice_status, hold_code, receipt_status, po_status, contract_status, price_type, supplier_risk_rating, *_action/*_actor must be the EXACT string from the API record. Do NOT derive or "clean up" a status (e.g. do not turn a receipt's literal accepted into partial; that loses points). If the template asks for a status that is literally a record field, echo the record field.
3. Enums are closed sets — copy them character-for-character from the template's allowed list. Never invent a value or change case/underscores.
4. USD amounts rounded to cents (2 decimals). Ratios/percentages to the precision the template states (e.g. completion ratio to 4, variance_pct to 1). Compute on raw numbers, round only the final field.
5. List fields are SETS unless the template states a sort rule. When a sort is stated ("sorted ascending", "by po_line_id ascending", "invoice_id ascending"), sort exactly that way. When it says "set; evaluator sorts", order doesn't matter but membership does — include every required element and nothing extra.
6. Null vs empty: use null when a value is genuinely absent (e.g. no contract → commercial_basis_id: null; no hold → hold_code: null), and [] for an empty list. Use 0.00 for a money/qty field that the template says defaults to zero (e.g. quantity_received when no receipt exists).
7. as_of / review_as_of date filtering: include only records dated on or before the as-of date. A later-dated receipt/invoice is OUT of scope even if it exists. Apply this before building any evidence/exception list.

---

3. Business rules (transferable)

3a. Three-way match (PO ↔ receipt ↔ invoice)

• Match on po_id + po_line_id + sku. Compare quantity (PO ordered), quantity_received (receipt), quantity_billed (invoice), and unit prices (PO vs contract vs invoice).
• short_qty_vs_po = ordered - received. unreceived_billed = max(billed - received, 0).
• receipt_completion_ratio = received / ordered.
• A clean three-way match = billed == received == ordered AND prices match AND a receipt exists AND invoice not on hold → release / APPROVED_THREE_WAY_MATCH.

3b. Quantity / variance reconciliation

• quantity_variance = billed - received. quantity_variance_pct = variance / PO_qty.
• If NO receipt exists, received = 0 and the variance equals the billed qty, but the controlling exception is NO_RECEIPT, NOT a quantity-variance code. Do NOT also stack a QTY_VARIANCE reason on a no-receipt invoice — the missing receipt supersedes it. (Conversely, a receipted-but-short invoice IS a qty variance.)

3c. AP invoice release / hold logic

• RELEASE only when: invoice status is approved (or otherwise clear), a valid receipt exists, quantities reconcile (within approved chargeback), and prices match. A scheduled payment found in /payments is a supporting RELEASE signal.
• HOLD when any of: status is on_hold / pending_receipt; hold_code is set; no receipt; unresolved quantity or price variance; receipt in inspection_hold; only a pending (not approved) chargeback exists.
• Reason/exception codes for an invoice = the union of the conditions that are actually true for it, chosen from the template's allowed set, then sorted/setted per the template. Add a code only when its condition holds.

3d. Duplicate-invoice / duplicate-receipt handling

• One PO can carry multiple receipts and multiple invoices. Tie an invoice to its own receipt_id. Other same-PO receipts that the invoice does not reference are excluded from this invoice's scope — list them in any excluded_same_po_receipt_ids-style field and flag a follow-up to hold the duplicate for its own/separate invoice. Do not net one invoice against another invoice's receipt.

3e. Contract price consistency

• contract_price_match = invoice unit_price == contract unit_price (for the SKU on an active contract). PO price should also equal contract price; a mismatch is a PRICE_VARIANCE/PRICE_MISMATCH exception. A SKU with no contract record → contract_unit_price absent and commercial_basis_id: null, blocker missing_contract.

3f. Contract & budget headroom (change-control / new buys)

• Program/budget headroom = budget_cap - committed_amount (from the program or the matching budget_snapshot; prefer the snapshot the task names).
• Contract ceiling usage = sum of subtotal of NON-cancelled POs on that contract (EXCLUDE status == cancelled). headroom_before = ceiling - noncancelled_subtotal.
• Requested exposure: contract/ceiling exposure = line subtotal (qty × unit_price, before tax & freight); budget exposure = subtotal + tax (add freight only if the memo supplies freight). Use the memo's tax rate.
• headroom_after = headroom_before - requested_subtotal; ceiling_ok = headroom_after >= 0.
• budget_after = remaining_budget - requested_total; budget_ok = budget_after >= 0.
• max_quantity_with_current_budget = floor(remaining_budget / per_unit_total) where per_unit_total = unit_price × (1 + tax_rate).
• Decision = combine the failing gates: budget-fail + approval-fail → hold_for_budget_and_approval; single failures map to the single-gate hold enum; contract SKU/price mismatch → reject_contract_mismatch; all gates pass → release_amendment. blocker_count = number of failing gates; ready_to_release = (blocker_count == 0).

3g. Approval gates (sourcing nomination & change-control)

• Find the latest approval_event for the source requisition (/approval_events?object_id=<REQ>), by event_date.
• approval_ok is TRUE only if the latest action is in the memo's "good actions" (typically just approved). A requisition whose latest action is submitted is NOT approved → approval_ok=false, action obtain_final_requisition_approval. (Requisition status: converted or approved on the requisition record is not the same as an approval event — follow whatever the task keys off.)

3h. Supplier-risk policy

• supplier_risk_rating = supplier.risk_rating verbatim. A rating of watch is a soft signal/blocker (supplier_watch), NOT an automatic hard hold — it's "context only" unless an OPEN SEVERE event exists.
• "Open" risk events = status in {open, monitoring}; exclude closed. Filter by supplier_id and by as-of date.
• "Severe" = severity high or critical. supplier_risk_ok is FALSE only when a severe OPEN event exists; medium/low open events are context (still listed in open_event_ids, but not in severe_open_event_ids, and don't fail the gate).

3i. Sourcing nomination gates & readiness

• Per line, collect: selected supplier, primary requisition, commercial basis (active contract id or null), package PO ids, receipt evidence (≤ as-of), invoice exceptions (on-hold/exception invoices ≤ as-of), open/monitoring risk events (≤ as-of), and blocker codes.
• Blocker codes (sorted ascending, from the allowed set): missing_contract, supplier_watch, open_supplier_risk, ap_hold, pending_receipt, late_due_date; none only when there are zero blockers.
• Readiness/decision mapping that scored best:
  • A line with a valid contract whose only blockers are clearable/soft (supplier_watch, an open non-severe risk event, an AP hold pending variance clearance) → at_risk + conditional_nomination.
  • A line with a hard blocker (missing_contract, pending_receipt / no receipt, no commercial basis) → not_ready + hold.
  • A clean line → ready + nominate.
• Roll-up: overall_readiness is the WORST line state — if ANY line is not_ready/hold, the program is not_ready and send_to_committee = "no". Do NOT upgrade the program to at_risk or set committee="yes" just because one line is conditional while another is held. Committee buckets (nominate_now/conditional/hold supplier ids) follow each line's decision.

3j. Chargeback netting (receiving/AP release file)

• Approved chargeback → net it off the invoice: net_release_amount = invoice_total - approved_chargeback_amount, decision release_net_after_approved_chargeback. Chargeback amount = basis_quantity × unit_cost from the register.
• Pending (e.g. pending_quality_review) chargeback or a receipt in inspection_hold → HOLD (hold_pending_quality_chargeback / inspection_hold_pending_chargeback); net release 0; the amount goes to pending_chargeback_amount, not approved.
• No receipt on the PO → hold_missing_receipt / no_receipt_on_po; net 0.
• primary_reason should reflect the chargeback's reason_code (e.g. "Underage Quantity" → approved_qty_chargeback; "AP Quantity Variance" → approved_ap_quantity_variance).
• Receiving-exception codes per receipt come from the register reason_code plus the receipt status (e.g. inspection_hold → "Inspection Hold").
• Totals: approved_chargeback_total/pending_chargeback_total = sums of the respective amounts; net_release_total = sum of net amounts of RELEASED invoices only. Release/hold invoice-id lists follow each decision.
• Sources: API record families are authoritative (procureops_po_records, procureops_receipt_records, procureops_ap_records) plus the local chargeback register; request notes and stale-alias notes are supporting-only.

3k. Vendor balance / AP close

• Per supplier: close_balance = opening_balance + invoice_total - scheduled_payments (use the opening balance the memo states, often 0.00). A payment scheduled through the close window reduces the balance.
• balance_status: FULLY_SCHEDULED when scheduled covers the invoices and none held; OPEN_HELD when any invoice is on hold; else OPEN_APPROVED.
• Program summary aggregates invoice_total / held_total / released_total / net_close_balance per program. Hold/release queues = invoice ids by decision, sorted ascending. total_close_balance = sum of supplier close balances.

---

4. Step-by-step SOP for a new task

1. Read prompt.txt and every payload file. Extract: target IDs / watch set, the as-of date, tax rate, opening balances, chargeback register, and any business-control notes (what to exclude, what counts as a "good" approval).
2. Open answer_template.json and list every required key, its type, enum sets, rounding precision, and sort/set rules. This is your output contract.
3. Pull the authoritative records from the API for each target ID, then fan out: for each PO get its receipts (?po_id=) and invoices (?po_id=); for each invoice get payments (?invoice_id=); for the supplier get risk events (?supplier_id=); for the contract get POs (?contract_id=); plus program/budget/approval records as needed.
4. Apply the as-of/scope filter and the relevant rules from section 3.
5. Compute money fields on raw numbers; round only at the end. Sort/ set lists per the template.
6. Assemble the JSON. Echo literal record statuses verbatim; copy enums exactly; use null/[]/0.00 per section 2.6.
7. Self-check before finalizing:
   • Every required key present, no extras, correct nesting.
   • Enums spelled exactly as allowed; no invented values.
   • Lists sorted where required; sets complete and minimal.
   • Money to cents; ratios/percentages to stated precision.
   • Roll-ups reconcile (queues vs decisions, totals vs line sums, worst-case readiness, blocker_count vs failing gates).
   • No-receipt invoices use NO_RECEIPT (not a stacked qty-variance code).
   • Literal statuses copied, not derived.

---

5. Common misjudgments to avoid

• Deriving/renaming a literal status field instead of echoing the API value.
• Stacking a quantity-variance reason on a no-receipt invoice (use NO_RECEIPT only).
• Treating supplier_watch (or a medium/low open risk event) as a hard hold — it is conditional context, not an automatic block; only a severe OPEN event fails the supplier-risk gate.
• Upgrading program-level readiness / committee routing because one line is conditional while another is held — the worst line governs the roll-up (any hold ⇒ not_ready, send_to_committee=no).
• Including cancelled POs in contract-usage / ceiling math (always exclude them).
• Confusing budget exposure (subtotal + tax) with contract/ceiling exposure (subtotal only, before tax/freight).
• Counting a requisition record's status as an approval; follow the latest approval event action against the memo's allowed "good" actions.
• Including records dated after the as-of date in evidence/exception lists.
• Netting an invoice against another invoice's receipt, or against a PENDING (not approved) chargeback.
• Missing the duplicate same-PO receipt that belongs to a separate invoice.
• Adding extra IDs to a "set" list (membership must be exact) or forgetting to sort a list that specifies ordering.