OCR COMPLETENESS GATE / SYNTHETIC PACKET EDITION v1.0.0

For automation developers who need a small, labelled packet-rule regression set before approval/export. Customer-specific labelled samples remain essential. This pack supplements them; it does not implement OCR.

QUICK START (Python 3.10+, standard library only)
  python checker.py cases/01-missing-attachment-corrected/extraction.json
  python checker.py cases/01-missing-attachment-negative/extraction.json
BLOCK prints JSON and exits 1. PASS exits 0. Invalid schema exits 2.

CONTENTS
FREE SAMPLE: one negative/corrected pair; two packet PDFs; source JSON and HTML; companion UTF-8 extracted text; extraction JSON and exact expected decisions/reasons; manifest and checker. PDFs are digital test inputs, not a printable planner. Each PDF page is one document; source_page in extraction.json is one-based. In case 04 negative the empty documents list deliberately has no extracted page provenance; submitted_page_count still describes the input PDF.

RULES
Exactly one invoice and one delivery_note; nonempty extraction text; no duplicate normalized document text; equal nonempty supplier_id and order_id. Document classification and fields are supplied by your adapter. The checker trusts those fields and does not re-extract or authenticate them. Compare both decision and reasons. Do not substitute filename inference for document classification. The duplicate fixture includes an invoice twice, rather than a delivery note relabelled in metadata. Case 04 changes extraction only: the two source packets are identical.

INTEGRATION
First run supplied extraction.json to verify the harness. Then OCR packet.pdf using your own engine and map output to this JSON shape: submitted_page_count integer; documents array of objects with kind, document_id, supplier_id, order_id, text and source_page. Do not feed expected.json into your pipeline. Compare with expected.json, investigate classification and field differences, and test your own layouts before deployment. Extracted.txt is a normalized field-text fixture, not a verbatim PDF text dump. Source PDFs contain additional synthetic disclosure and layout labels.

The weak_baseline only counts submitted pages. It is deliberately defective and educational, not a competitive benchmark. No claim about any real OCR system or loss prevention follows from its misses. All outcomes are declared synthetic rules, not compliance judgments. No OCR engine, real client integration, physical print, human usability study or independent accuracy evaluation was performed. This tiny set does not test adversarial inputs, signatures, tables, degraded scans, multilingual layouts or broad production coverage. Treat client inputs as untrusted and validate schemas separately.

RIGHTS AND DISCLOSURE
Original AI-assisted fictional documents, code and labels. Invented IDs; no real invoices, people, logos, signatures, customer records or third-party benchmark files included. See REUSE.txt. No ownership/registration or exclusive copyright guarantee is asserted.

COMMERCIAL STATUS
Public free sample. Proposed full-pack price $12 USD is unvalidated; full-pack enquiries only. Orders and payment are closed. No paid demand, OCR-engine evaluation or payment readiness is represented. Support/fulfillment not promised by this artifact. Public sample: https://moxie.staticterminal.com/ocr-completeness-gate/
