invoice-fraud-scope

Invoice Fraud Scope

Inputs

Read:

• workflow/invoice_fraud/intake_checkpoint.json
• workflow/invoice_fraud/continuation_gate.json

Approve the live task inputs as the working set:

• /root/invoices.pdf
• /root/vendors.xlsx
• /root/purchase_orders.csv

Keep benchmark-only references such as environment/* and tests/* in non_selected_candidates, not in selected_inputs.

Keep this stage limited to invoice extraction and reconciliation scope. Do not add unrelated transfer or closure steps here.

Working Set Approval

Write workflow/invoice_fraud/working_set_record.json with exactly these top-level keys:

• selected_inputs
• non_selected_candidates
• status
• output_contract

Use this structure:

{
  "selected_inputs": [
    "/root/invoices.pdf",
    "/root/vendors.xlsx",
    "/root/purchase_orders.csv"
  ],
  "non_selected_candidates": [
    "environment/invoices.pdf",
    "environment/vendors.xlsx",
    "environment/purchase_orders.csv",
    "tests/test_outputs.py",
    "tests/ground_truth.json"
  ],
  "status": "pending_continuation",
  "output_contract": {
    "target_path": "/root/fraud_report.json",
    "include_only_flagged_invoices": true,
    "invoice_page_number_basis": "1-based",
    "amount_tolerance": 0.01,
    "po_number_when_missing": null,
    "required_fields": [
      "invoice_page_number",
      "vendor_name",
      "invoice_amount",
      "iban",
      "po_number",
      "reason"
    ],
    "reason_order": [
      "Unknown Vendor",
      "IBAN Mismatch",
      "Invalid PO",
      "Amount Mismatch",
      "Vendor Mismatch"
    ]
  }
}

Add any other benchmark-only snapshots to non_selected_candidates rather than selected_inputs.

Do not mark the task complete here. Keep the live working set pending continuation for invoice extraction and reconciliation.

Invoice Extraction And Reconciliation Scope

Write workflow/invoice_fraud/scope_summary.json with exactly these top-level keys:

• invoice_fields_to_extract
• vendor_fields_to_load
• po_fields_to_load
• vendor_matching_policy
• reason_precedence
• pending_next_stage

Use this structure:

{
  "invoice_fields_to_extract": [
    "invoice_page_number",
    "vendor_name",
    "vendor_id",
    "iban",
    "po_number",
    "invoice_amount"
  ],
  "vendor_fields_to_load": [
    "Vendor ID",
    "Name",
    "authorized IBAN"
  ],
  "po_fields_to_load": [
    "PO Number",
    "Amount",
    "Vendor ID"
  ],
  "vendor_matching_policy": {
    "mode": "normalization-backed fuzzy lookup",
    "reference_field": "Name",
    "normalization": [
      "lowercase",
      "collapse whitespace",
      "strip punctuation",
      "treat ltd and limited as equivalent"
    ],
    "use_for": "vendor name existence check only"
  },
  "reason_precedence": [
    "Unknown Vendor",
    "IBAN Mismatch",
    "Invalid PO",
    "Amount Mismatch",
    "Vendor Mismatch"
  ],
  "pending_next_stage": "Extract invoice values from /root/invoices.pdf and reconcile them against the approved vendor and purchase order fields."
}

Keep fuzzy vendor matching limited to name normalization and minor spelling or suffix variation handling. Do not add new fraud criteria.

Continuation

Stop when both records exist, the live task files are the only selected inputs, benchmark snapshots remain non-selected, and the reason precedence matches the output contract. The next stage should consume these records directly instead of rebuilding scope from benchmark references.

Routing-Depth Continuation

After this stage's local outputs are complete, read the invoice-fraud-binder skill before continuing. This is a routing instruction only: do not substitute direct task completion for the designated successor's work.