On this page
Define the page and text contract first
The source dates shown in this guide identify documentation checks on 22 October 2026, not original publication dates. Follow each citation for the document format vocabulary being discussed. Record a controlled document id, page id, source image dimensions, orientation, resolution if available, OCR engine or workflow version, and output format. Keep source images and raw OCR outside a public worksheet unless the owner has approved the handling boundary.
| Layer | Synthetic field | Question to resolve |
|---|---|---|
| Document identity | synthetic-doc-04 | Is this the source image or a derivative? |
| Page geometry | width 1200, height 1600 | Which coordinate units and origin apply? |
| Text unit | word, line, block, or region | What does one record represent? |
| Position | left, top, width, height | Does the box map to the same page? |
| Order | block 2 then block 1 | Is reading order explicit or inferred? |
| Confidence | synthetic score, scale unknown | What produced the score and scale? |
| Processing | workflow-3, checked date | Which activity generated this output? |
The Library of Congress, (opens in a new tab) describes layout and content for physical text resources. The Tesseract OCR, (opens in a new tab) documentation names output fields that can expose page, block, line, word, coordinates, confidence, and text. Treat those fields as observations from a particular workflow, not universal semantics for every OCR export.
Reconcile coordinates, boxes, and reading order
Coordinate review starts with page width, height, origin, units, rotation, and crop. Compare a word box to its parent line, block, and page. A box that fits page bounds can still point to the wrong page version or be misplaced after rotation. Record raw coordinates and derived checks separately.
Synthetic OCR row fixture
{
"document_id": "synthetic-doc-04",
"page": {"id": "p1", "width": 1200, "height": 1600, "origin": "top-left"},
"rows": [
{"level": "word", "id": "w1", "text": "alpha", "left": 80, "top": 120, "width": 90, "height": 30, "order": 1, "confidence": "synthetic"},
{"level": "word", "id": "w2", "text": "beta", "left": 185, "top": 120, "width": 72, "height": 30, "order": 2, "confidence": "synthetic"}
],
"review": {"page_link": "matched", "bounds": "inside", "reading_order": "conditional"}
}The text in this fixture is synthetic. A real report should avoid private source text and use stable row identifiers, aggregate counts, or owner-approved redactions. If page rotation or image preprocessing occurred, include that activity in the transform record rather than treating coordinates as timeless.
| Finding | Evidence statement | Disposition |
|---|---|---|
| Box outside page | Coordinates exceed the recorded page geometry | Hold page or transform link |
| Text and box disagree | Text row is linked to a different region | Retain both records and investigate |
| Order not stated | Coordinates exist but sequence is absent | Mark reading order unknown |
| Confidence scale absent | A score is present without scale or policy | Report value as uncalibrated evidence |
| Page ids differ | OCR and image references do not reconcile | Hold artifact identity |
Separate text fidelity from layout evidence
A text match can be correct while a coordinate is wrong; a coordinate can align while a character is misread. Keep word, line, block, and page findings separate. Record an edit or normalization as a new derived text entity with the reason and source row references. Do not silently replace the OCR output.
| Check | Record | Do not infer |
|---|---|---|
| Character finding | Source row id and normalized observation | That surrounding words are correct |
| Word boundary | Token span and neighboring boxes | That punctuation was preserved |
| Line grouping | Parent line and order evidence | That reading order is natural |
| Region type | Header, body, table, or unknown | That a layout label is semantic truth |
| Confidence | Value, scale, and producing workflow | That confidence is calibrated |
| Correction | New text version and reason | That edited text is source text |
- Check page identity and geometry before interpreting word rows.
- Compare parent-child boxes and order using synthetic fixtures.
- Classify text, layout, and confidence findings separately.
- Keep corrections linked to source rows and workflow versions.
- Publish only controlled excerpts or aggregate diagnostics.
Use a worksheet for layout failures and unknowns
Use one row per page or region in the review ledger. Include source references for page images, OCR output, preprocessing activity, and any correction. Keep unknown values explicit so an absent confidence scale is not mistaken for a failed confidence check.
| Review field | Example state | Next question |
|---|---|---|
| Page identity | matched / mismatch / unknown | Which image version is authoritative? |
| Dimensions | matched / transformed / unknown | What operation changed geometry? |
| Coordinates | inside / outside / not supplied | Do boxes use the stated units? |
| Hierarchy | word to line to block | Are parent relationships supported? |
| Reading order | explicit / inferred / unknown | Who owns a disputed order? |
| Text correction | none / derived version | What source row supports the edit? |
| Confidence | value with scale / unavailable | Can it be compared at all? |
| Disposition | conditional / hold / unresolved | What evidence is still required? |
Keep OCR evidence scoped and qualified
ALTO and Tesseract documentation describe output structures and fields, not the correctness or authority of a particular document. This review can show which page, region, coordinate, text version, and processing record were supplied. It cannot establish that text is accurate, complete, accessible, private, authorized, or suitable for a downstream use.
Pair this guide with the data provenance log for processing history and annotation label ontology versioning when layout labels have versioned meanings. After supported non-health scope, authority to describe the collection, and a controlled OCR packet are established, the qualified appraisal request path may be a next preparation step. It does not run OCR, correct text, create a quote, match a buyer, issue a legal opinion, or grant a license.
Takeaways
- Record document and page identity, dimensions, coordinate convention, hierarchy, order, confidence, and processing activity.
- Keep text, layout, and confidence findings separate.
- Retain source OCR and represent corrections as linked derived records.
- Use synthetic rows to test page, box, order, and transform reconciliation.
- Treat format fields as evidence, not a universal accuracy or rights conclusion.