Self Attempt 02

ProcureOps procure-to-pay solver skill

You answer procurement / receiving / accounts-payable questions against the ProcureOps read-only ERP API and a per-task local payload. Output is always a single JSON object that must match the task's answer_template.json shape exactly.

0. Golden rules (read first)

1. The API is the source of truth. Local payloads (memos, packets, chargeback registers, export notes) only tell you the business request and the watch set (which POs/receipts/invoices/suppliers to look at). Whenever a number disagrees, trust the API record, not the payload. One documented exception: a local chargeback / credit register is authoritative for chargeback amounts and approval status because those debits do not live in the ERP yet (see Task-5 pattern).
2. Base URL is supplied by the runner / environment_access.md. The prompt's http://127.0.0.1:8006 or localhost:8006 is a placeholder — replace it with the real base URL from environment_access.md (e.g. <remote-env-url>).
3. Match the template byte-for-byte. Same top-level keys, same nested keys, same enum spellings (copy enum strings verbatim, including UPPER_SNAKE vs lower_snake). Do not add keys the template does not list. Read whether task_id is a literal required_value and emit exactly that string.
4. Rounding & types. USD amounts → round to 2 decimals (cents). Ratios with a stated precision (e.g. precision: 4) → round to that many places. Quantities are integers. Percentages → the stated decimals (often 1). Booleans are real JSON booleans, not strings.
5. Lists are sets unless a sort rule is given. If the template says "sorted ascending" / "ordering: X ascending", sort. If it says "set; evaluator sorts values", order does not matter but de-duplicate. When in doubt, sort ascending — it never hurts a set.
6. Respect the as-of / close / review date. Records dated after the as-of date are out of scope. The date filter is inclusive of the as-of day (a record dated exactly on the as-of date is IN scope). This is the single most common scoping mistake — see §5.

1. API usage

Endpoints (GET only): GET / (endpoint list), GET /health, GET /manifest (counts + anchor IDs), GET /<collection> → {"count": N, "results": [...]}, GET /<collection>/<id> → one object (404 if missing).

Collections: programs, suppliers, items, contracts, purchase_requisitions, purchase_orders, receipts, ap_invoices, payments, approval_events, budget_snapshots, vendor_risk_events. Aliases also work: purchase-orders, ap-invoices / /ap/invoices, /ap/payments, approvals, budgets, vendor-risks, purchase-requests.

Filtering: append any record field as a query param (case-insensitive), e.g. /receipts?po_id=PO-AX17-4481, /ap_invoices?supplier_id=SUP-LUMA, /vendor_risk_events?supplier_id=SUP-LUMA, /approval_events?object_id=REQ-AX17-141, /purchase_orders?contract_id=CR-LMP-228, /purchase_orders?sku=LMP-228. Date window via start/end on the collection's primary date field (receipts→receipt_date, ap_invoices→invoice_date, payments→scheduled_date, purchase_orders→order_date, approval_events→event_date, budget_snapshots→snapshot_date, vendor_risk_events→event_date, purchase_requisitions→need_by, contracts→effective_date). end=<as_of> is the clean way to scope to the as-of date and it is inclusive.

Workflow: pull the named anchors by ID, then fan out by filter (all receipts for a PO, all invoices for a PO, all risk events for a supplier, all POs for a contract). Never assume one PO has one receipt or one invoice — multiples are common and scoping them correctly is where points are won/lost.

2. Record schemas (fields you will use)

• programs: program_id, name, owner, cost_center, budget_cap, committed_amount, priority, status, region.
• suppliers: supplier_id, name, risk_rating (low|watch|medium|high), status (active|quality_hold), payment_terms, region.
• items: sku, description, category, standard_cost, preferred_supplier_id, uom, active.
• contracts: contract_id, program_id, sku, supplier_id, status (active|draft|...), price_type (fixed|...), unit_price, ceiling_amount, effective_date, expiry_date, buyer.
• purchase_requisitions: requisition_id, program_id, sku, quantity, requester, need_by, priority, status (approved|converted|...).
• purchase_orders: po_id, program_id, supplier_id, contract_id (may be null), requisition_id, status (open|confirmed|partial_receipt|received|closed|cancelled), order_date, due_date, ship_to, currency, subtotal, tax, total, lines:[{line_id, sku, description, quantity, unit_price}].
• receipts: receipt_id, po_id, supplier_id, warehouse_id, receipt_date, packing_slip, receiver, status (accepted|accepted_with_note|inspection_hold), lines:[{po_line_id, sku, quantity_received, quantity_rejected, inspection_status (passed|variance)}].
• ap_invoices: invoice_id, po_id, supplier_id, receipt_id (may be null), invoice_date, status (approved|paid|on_hold|pending_receipt|entered), hold_code (null|NO_RECEIPT|PRICE_VARIANCE|QTY_VARIANCE|SUPPLIER_REVIEW), currency, subtotal, freight, tax, total, lines:[{po_line_id, sku, quantity_billed, unit_price}].
• payments: payment_id, invoice_id, supplier_id, amount, scheduled_date, status (scheduled|...), currency.
• approval_events: event_id, object_id (e.g. a requisition_id), object_type, action (submitted|approved|escalated|returned), actor, event_date, note_code.
• budget_snapshots: snapshot_id, program_id, budget_cap, committed_amount, pending_invoice_amount, snapshot_date, currency.
• vendor_risk_events: event_id, supplier_id, event_type (bank_change|invoice_variance|late_delivery|duplicate_invoice_review|quality_hold), severity (low|medium|high), status (open|monitoring|closed), related_object_id, event_date.

3. Core business rules (transferable)

3.1 Three-way match (PO ↔ receipt ↔ invoice)

Match on the same po_line_id/sku. Compute per line:

• short_qty_vs_po = ordered_qty - received_qty (PO line qty minus receipt qty).
• unreceived_billed_qty = billed_qty - received_qty (positive = invoiced more than received → the dangerous case).
• receipt_completion_ratio = received_qty / ordered_qty.
• quantity_variance = billed_qty - received_qty; quantity_variance_pct = variance / PO_line_quantity * 100 (variance is taken as a % of the PO quantity, not of billed — read the template wording). A clean three-way match = invoice approved, no hold_code, billed == received == ordered (or billed <= received), and prices consistent.

3.2 Quantity_received: which receipts count

• For an invoice-level reconciliation, the invoice usually carries a receipt_id. Use that specific receipt's received quantity. If receipt_id is null → received = 0 and the invoice is effectively NO_RECEIPT.
• For a PO/receiving-batch reconciliation, scope to the named batch/receipt only (the memo will say "closeout for an already-posted receipt"). Do NOT roll in later receipts on the same PO that fall after the as-of/review date.
• Always check whether a PO has other receipts: an extra receipt on the same PO that is out of the date window or tied to a different invoice goes into an excluded list (e.g. excluded_same_po_receipt_ids), not the in-scope list.

3.3 Contract price consistency

contract_price_match (or ceiling_ok pricing context) = invoice/PO unit_price equals the active contract's unit_price for that sku/supplier. A mismatch → PRICE_MISMATCH / PRICE_VARIANCE. Find the contract via /contracts?sku=...&supplier_id=...; if none exists the line has no commercial basis (blocker missing_contract, commercial_basis_id = null).

3.4 Contract ceiling headroom

• noncancelled_subtotal = sum of subtotal over all POs on the contract whose status is not cancelled (exclude cancelled; list them in excluded_cancelled_po_ids).
• headroom_before_change = ceiling_amount - noncancelled_subtotal.
• requested_subtotal = requested_quantity * contract_unit_price.
• headroom_after_change = headroom_before_change - requested_subtotal.
• ceiling_ok = requested_subtotal <= headroom_before_change (equivalently headroom_after >= 0).
• Contract ceiling exposure is subtotal before tax and freight.

3.5 Program budget headroom

• Use the program's budget_snapshot (preferred) or the program record: remaining_budget = budget_cap - committed_amount.
• Budget exposure = line subtotal + estimated tax (freight only if the memo supplies a freight figure). requested_tax = requested_subtotal * tax_rate; requested_total = requested_subtotal + requested_tax.
• budget_after_change = remaining_budget - requested_total; budget_ok = budget_after_change >= 0.
• max_quantity_with_current_budget = floor( remaining_budget / (unit_price * (1 + tax_rate)) ) (per-unit cost grossed up for tax; floor it).
• budget_headroom_usd for a program summary = budget_cap - committed_amount (rounded to cents).

3.6 Approval gate

Find the requisition's events via /approval_events?object_id=<requisition_id>. Take the latest by event_date (tie-break by event_id). approval_ok is true only if the latest action is in the memo's "good actions" set — typically just approved. submitted, escalated, returned do not clear the gate.

3.7 Supplier-risk policy

• risk_rating of watch/medium/high is context only and does not by itself block a release — unless the rule says so. A watch rating commonly raises a supplier_watch blocker code or a SUPPLIER_WATCH_RISK exception note in nomination/receiving contexts, but it does not force a HOLD on its own.
• Open risk events = events with status open or monitoring (closed are out). Filter by /vendor_risk_events?supplier_id=... then keep open/monitoring as of the as-of date (event_date <= as_of).
• Severe open event = an open/monitoring event with severity high (treat critical likewise if present). A severe open event is a real blocker (open_supplier_risk / hold_for_supplier_risk); supplier_risk_ok = false. Medium/low open events are usually monitored, not blocking — unless the template's logic counts any open event.
• Supplier status = quality_hold (vs active) is itself a hard supplier block.

3.8 AP invoice release vs hold

Default to HOLD / keep_on_hold when ANY of:

• invoice status is on_hold or pending_receipt, or any non-null hold_code;
• hold_code = NO_RECEIPT or the invoice's receipt is missing (received = 0);
• billed_qty > received_qty (INVOICE_QTY_EXCEEDS_RECEIPT / QTY_VARIANCE);
• price mismatch vs contract (PRICE_VARIANCE);
• a severe open supplier-risk event / supplier on quality_hold. RELEASE only when the invoice is approved with no hold_code, the three-way match is clean (billed == received), and there is no blocking risk. A scheduled payment in /payments?invoice_id=... (with scheduled_date within the horizon) is evidence to release and adds SCHEDULED_PAYMENT_FOUND.

3.9 Duplicate-invoice handling

Two non-cancelled invoices on the same PO (especially similar totals, or a duplicate_invoice_review risk event) signal a possible duplicate. Do not blindly release both — keep the later/unconfirmed one on hold and flag it. When the task scope names ONE invoice per PO, the other invoice/receipt on that PO is excluded context, not in-scope.

3.10 Chargeback / credit netting (AP release with debits)

When a local chargeback register lists debits against an invoice:

• approved chargeback → net it now: net_release_amount = invoice_total - approved_chargeback_amount; decision = release_net_after_approved_chargeback.
• pending_* chargeback (e.g. pending_quality_review) → do NOT net, HOLD the invoice: decision = hold_pending_quality_chargeback, net_release = 0, pending_chargeback_amount = the pending debit.
• chargeback amount = basis_quantity * unit_cost (round to cents).
• An invoice with no receipt on its PO → hold_missing_receipt, net_release = 0, regardless of any release request note.
• A receipt in inspection_hold with a pending chargeback → hold for quality.

4. Output conventions

• Build the JSON to mirror the template's nesting and key order exactly; emit only the keys the template names. If the template is "shape doc" (with type, required_keys, allowed_values), produce the concrete object those describe, not the meta-schema.
• Enums: copy the allowed string verbatim. Reason-code / exception-code lists: include only codes that actually apply; if none apply and a NO_EXCEPTION / none sentinel exists, use it.
• Sort every list the template marks "ascending" / "sorted"; de-dup set lists.
• USD → 2 dp; ratios → stated precision; integers for quantities; real booleans.
• For "queue" outputs (hold queue / release queue), list invoice_ids and sort ascending; an invoice appears in exactly one queue.
• Evidence/supporting-id lists: include the actual record IDs you read from the API; keep authoritative vs supporting-only sources separate when the template asks (API records = authoritative; local notes/aliases = supporting-only).

5. Common misjudgments (learned from train tasks)

1. Including out-of-window receipts/invoices. A PO often has a later receipt (e.g. dated 06-08 when as-of is 06-01) — exclude it. Use end=<as_of> or filter date <= as_of yourself. Boundary date == as-of is INCLUDED.
2. Rolling all receipts on a PO into one invoice's received qty. Use the invoice's own receipt_id; treat the rest as excluded/duplicate context.
3. Treating watch/medium risk as an automatic HOLD. It is context only; only a severe (high) open event, supplier quality_hold status, or an actual match/qty/price failure blocks a release.
4. Counting cancelled POs in contract usage. Always exclude cancelled from noncancelled_subtotal and list them separately.
5. Approval false-positives. "submitted"/"escalated"/"returned" are NOT approvals; only the configured good action (usually approved) clears the gate, and you must use the latest event.
6. Variance % base. Variance percentage is of the PO quantity, not billed quantity — read the field description.
7. Tax in budget but not in ceiling. Contract ceiling exposure = subtotal only; program budget exposure = subtotal + tax (+ freight only if memo gives it).
8. task_id / required_value. Emit the literal value the template demands (e.g. "train_002"), not the API-style id.
9. Local payload numbers vs API. If the memo and API disagree on a price/qty, use the API — except for chargeback debits that live only in the local register.

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

1. Read prompt.txt, the local payload(s), and answer_template.json. Note: the as-of/close/review date, the exact watch-set IDs, the tax rate, any "good actions" / exclusion rules, and the literal task_id/required values.
2. Map the template: list every key, its type, enum domain, rounding, and sort/set rule. This is your output contract.
3. Pull each named anchor by ID from the API. Then fan out by filter to get every related receipt (per PO), invoice (per PO/supplier/program), payment (per invoice), approval event (per requisition), risk event (per supplier), and contract (per sku+supplier or contract_id).
4. Apply the as-of date scope: drop records dated after the as-of date; keep the exact watch-set; separate excluded/duplicate context records.
5. Run the relevant rule set (§3): three-way match, price/contract check, ceiling headroom, budget headroom, approval gate, supplier-risk, AP hold/release, chargeback netting. Compute every numeric field with explicit rounding.
6. Derive decisions/enums from the computed facts; assemble queues and reason/ exception code lists (apply, don't over-add).
7. Emit JSON matching the template exactly — right keys, enum spellings, rounding, sorted/de-duped lists, real booleans. No prose outside the JSON object.
8. Self-check: keys present == template keys; every amount 2 dp; every "ascending" list sorted; every set de-duped; literal task_id correct; decision consistent with the blocker/reason codes you emitted (e.g. if any blocker, ready_to_release is false and blocker_count matches the count of required actions).