Eleven pages that break OCR and document AI pipelines, one from each of our open synthetic benchmarks. Every page is generated, so the answer key is exact and there's no customer data anywhere.
For each one: what the page looks like, why pipelines fail on it, and an excerpt of the answer key.
A rotated APPROVED stamp sits on top of the bill-to name and address. OCR merges stamp ink with the text underneath, so the name comes out garbled, or the stamp text ends up inside the field.
| bill_to_name | Juniper Healthcare Systems Corp. |
| bill_to_address | 5984 Station Lane, Calgary, AB T2P 4Q6, Canada |
| total_amount | CA$3,813.33 |
HyperAPI extract returned all 10 fields and all 5 line items correctly on this page, including both fields under the stamp.
Twelve fields filled in by hand on a printed claim form, plus margin notes ("verified ID", "see attached") that belong to no field. Digits like 1/7 and 0/6 are where extraction usually slips.
| claim_number | CLM-060721934 |
| policy_number | POL-56941089 |
| date_of_birth | 10/13/1982 |
Severe degradation: scaled to about 51 effective DPI, blurred, noisy, low contrast, heavy JPEG compression, around 2,000 dust specks and a stamp on top. The answer key is unchanged from the clean page.
| po_number | PO-57717 |
| date | 14/02/2025 |
| supplier_name | Cedar Lab Supplies Co. |
Three dense columns of abbreviated listings in small serif type on aged paper. Read straight across the page and every listing merges with its neighbours, with no error to warn you.
| kind | directory_listing |
| reading order | column by column, top to bottom |
| labels | every block with its bbox and order |
Nine PII spans of six types (name, date of birth, address, phone, email, account number) spread across a patient intake form. A redactor that misses one leaves the page non-compliant, so the metric that matters is the leak rate.
| PII spans | 9 |
| types | PERSON_NAME, DATE_OF_BIRTH, ADDRESS, PHONE, EMAIL, ACCOUNT_NUMBER |
| boxes | pixel boxes for every span |
Five tables and 17 layout blocks on one page, with comparative-year columns. Parsers drop rows, merge adjacent tables or lose the column alignment between years.
| tables | 5, each with HTML ground truth |
| blocks | 17, with reading order |
| scoring | CER, TEDS and Kendall tau |
The page is titled "Aging Statement", but its class is statement of account, and the scan is degraded. A classifier that keys on the title gets it wrong.
| label | statement_of_account |
| title_on_page | Aging Statement |
| variant | degraded |