# PDF extraction review worksheet

OpenMax editorial worksheet · revised September 4, 2026. Copy this file for each selected document. This is a requirements and review aid, not a verified product integration or authorization to upload, pay or write. Do not put confidential documents into an unapproved service. The nine sections contain 45 labeled fields; leave unknown facts explicitly unresolved.

## 1. Received file and access

- Selected original or controlled reference:
- Received time, timezone, byte size and actual file hash:
- Permitted users, purpose and retention rule:
- Approved intake/parser environment and relevant limits:
- Derived files and their links to the original:

A hash identifies bytes, not the issuer or authenticity. Keep the original distinct from an OCR or redacted copy.

## 2. Document family and selected version

- Supported document family and intended output:
- Document business key and issuer verification handled separately:
- Selected revision and evidence for selection:
- Amendments, replacements and excluded versions:
- Classification uncertainty, assigned reviewer and resolution:

Do not combine a revised cover with outdated item pages without a documented basis.

## 3. Page inventory

- Actual PDF page count and expected sections:
- Mapping of zero-based page indexes to printed labels:
- Continuation statements and attachment relationships:
- Missing, duplicated, rotated or cropped pages:
- Page-completeness decision and unresolved reasons:

An empty extracted string does not establish a blank page. In the teaching PDF, page index 2 contains the last item and totals as an image.

## 4. Extraction method

- Text, image or mixed classification by required page/region:
- Chosen parser/OCR/manual path and exact tool version:
- Verified input languages, size constraints and assumptions:
- Reading-order or visible-versus-hidden-text conflicts:
- Failed extraction handling and approved fallback:

Record methods actually used. Do not turn a planned OCR call into a claimed test result.

## 5. Field evidence and geometry

- Per-field original page and raw visible value:
- Field box/polygon, units, origin and coordinate convention:
- Rendering size, rotation and crop transformation:
- Item-row and column associations across pages:
- Relevant captions, units and footnotes outside table bounds:

Keep the actual extraction geometry separate from the sample's authored rectangles. A highlighted box alone does not prove the right field was selected.

## 6. Types and normalization

- Field schema version, names, types and cardinality:
- Date/number locale rule and supporting source statement:
- Raw values, proposed normalized values and transformations:
- Missing, unreadable, not_set or not-applicable states:
- Type/format conflicts and required clarifications:

Preserve identifier characters. In the exercise Q-00742 stays a string, 03/04/2026 follows DD/MM/YYYY, and an unset delivery date remains null/not_set.

## 7. Cross-field validation

- Supported calculations and relationship checks:
- Exact decimal and rounding rules:
- Input field references and recomputed results:
- Differences, affected fields and tolerance justification:
- Resolution evidence without silently overwriting the source:

For the exercise, 300 + 300 + 240 = 840; 840 + 168 = 1008 EUR; candidate 100.80 differs by −907.20 EUR. Tax is a fictional teaching assumption.

## 8. Duplicate candidates and review signals

- Exact-byte duplicate result and associated file references:
- Business-key/version duplicate candidates:
- Confidence values, provenance and field-specific review rule:
- Structural anomalies needing investigation, without fraud conclusions:
- Review priority, responsible person and pending questions:

Confidence is not measured accuracy. Metadata anomalies are not evidence of fraud by themselves.

## 9. Record release and destination

- Accepted/corrected/rejected fields, reviewer and reason:
- Record-level release decision and unresolved blockers:
- Authorized destination, mapping and draft/production boundary:
- Existing-record check, retry identification and write outcome:
- Correction path and links to earlier released versions:

Keep field acceptance, record release and write authorization separate. In the candidate exercise, 9/10 nonempty and 6/10 matching fields do not release the record: its four important errors remain blocked. This file does not grant production access.

## Review outcome

Document and version: ______. Confirmed by: ______. Review date/time: ______. Outstanding requirements: ______. Next permitted action: ______.

Obtain appropriate security, privacy, legal or financial review for consequential use. The worksheet is not a security certification, document-authenticity check or supplier/payment approval.