P05 The Paperwork Trial Generated invoice images
PO Revision
A generated scan case where split payments and the latest purchase-order revision decide the outcome.
Transparency boundary
This page shows the files the model can inspect. It does not show the hidden oracle solution.
Excluded: ground_truth.json, expected_artifacts.json, manual readings, and calibration notes.
Generated Images
Scans and visual attachments visible to the model
Task And Source Files
No oracle solution filesREADME_TASK.md
# Paperwork Trial v3 Generated Image Case 05 This is synthetic benchmark bookkeeping. It is not financial, tax, or legal advice. Audit the scanned document folder and write `audit_result.json` only. Case ID: `P3-GEN-05` Output keys exactly: `case_id, approved_invoice_ids, review_invoice_ids, reject_invoice_ids, ignored_document_ids, total_approved_gross_cents, warnings_by_invoice, evidence, proof_code` Rules: - Treat the files in `scans/` as the source documents. - Ignore documents that are explicitly quotes or not invoice payment requests. - Match vendors by visible vendor name and tax ID when available. - `payment_match` requires paid bank rows for the invoice whose summed amount equals the invoice gross. - Split bank payments count as `payment_match` when paid rows with the same invoice ID sum to the invoice gross. - Use the latest purchase order revision visible in the scanned documents and reflected in `purchase_orders.csv`. - `po_cancelled` applies when the invoice uses a purchase order that the latest revision marks cancelled. - Approved invoices have no warnings and paid amount equals gross. - Review invoices have warnings but are not reject-level. `po_cancelled` is review-level. - Reject only if the invoice is impossible to process, uses an inactive vendor, or lacks a required PO. - Sort all invoice-id arrays ascending. - Allowed warning codes are exactly: `po_cancelled`. - `warnings_by_invoice` must include every real invoice ID and sorted warning arrays. - Warning arrays must be flat arrays of lowercase strings, never nested arrays and never prose labels. - `ignored_document_ids` must include visible document IDs from ignored non-invoice documents, not filenames. - `total_approved_gross_cents` is the sum of approved invoice gross totals only. - `evidence` must list the relative source file paths used in stable alphabetical order, including folder prefixes such as `scans/`. - Include the relevant CSV files and every scanned document inspected in `evidence`, including ignored quote scans. - `proof_code = total_approved_gross_cents + sum(numeric parts of all real invoice IDs) + 97 * total_warning_count`.
bank_export.csv
date,description,invoice_id,amount_cents,status 2026-07-02,Delta Repair Group partial payment,INV-5600,20000,paid 2026-07-03,Delta Repair Group final payment,INV-5600,10000,paid 2026-07-04,Delta Repair Group parts invoice,INV-5601,18000,paid
model_prompt.md
You are auditing a synthetic scanned paperwork folder. Read `README_TASK.md`, inspect the files in `scans/`, and use `bank_export.csv`, `vendor_master.csv`, and `purchase_orders.csv`. Write `audit_result.json` only. This is benchmark bookkeeping, not financial, tax, or legal advice. Important: - The scanned images are the source documents. - Ignore quotes and documents that are not invoice payment requests. - Split payments can sum to one invoice. - Use the latest purchase order revision visible in the scanned documents and in `purchase_orders.csv`. - Do not invent fields that are not supported by the files. - Use visible document IDs for `ignored_document_ids`, not filenames. - Use relative paths with folder prefixes in `evidence`, for example `scans/example.png`. - Use only allowed lowercase warning codes from `README_TASK.md`. - Warning arrays must be flat arrays of strings.
purchase_orders.csv
po_id,vendor_id,limit_cents,status PO-5600-A,V-640,35000,open PO-5600-B,V-640,22000,cancelled
vendor_master.csv
vendor_id,name,tax_id,status V-640,Delta Repair Group,DR-640,active V-641,Delta Repair North,DR-641,active