INDUSTRY FIELD GUIDE
Document preparation and accessible reading-order data for AI
A document can look correct on screen while extraction interleaves columns and assigns text to the wrong prompt.
NAICS 561410 · Document Preparation Services
A document can look correct on screen while extraction interleaves columns and assigns text to the wrong prompt. W3C documents a PDF technique for reading and tab order. Document specialists can build original examples to evaluate whether a model reads the intended sequence and preserves each field’s label. W3C PDF reading-order technique
Design a fictional two-column application with a note that belongs to just one field. Export a correctly ordered version and a visually identical version with a misplaced note in the tag tree. Record the intended order and ask reviewers to verify both the extracted output and the source tags. Treat a correct answer from visual inference as a separate result from correct extraction.
Use text and graphics your team created or can license for the purpose. Leave client forms, signatures and document metadata out of the package.
Describe an original document preparation example
Original Fiund exercise
A visually correct two-column page can associate a note with the wrong extracted field
An original document-design fixture is described through intended and extracted sequences. No actual PDF export is claimed in this text packet; it specifies a reproducible evaluation to construct with owned content.
| Record | Source | What it records |
|---|---|---|
| RO1 | Intended layout | Left field A: preferred delivery day, note 'weekdays only.' Right field B: box color, no note. Original intended order A, A-note, B. |
| RO2 | Correct extraction fixture | Sequence: A prompt; weekdays-only note; B prompt. Tag-parent map places note under A. |
| RO3 | Misordered fixture | Same visual arrangement, extracted sequence A prompt; B prompt; weekdays-only note. Tag-parent map incorrectly places note under B. |
Question for the model
Which fixture preserves the note's intended field, and why should a visually inferred correct answer be scored separately from correct extraction?
Show the reference answer and checks
RO1 and RO2 attach the weekdays-only note to delivery-day field A. RO3's extraction and tag-parent map attach it after and under box-color field B, violating the intended association despite an unchanged visual arrangement. A model might infer the intended meaning from the rendered page, but that would not demonstrate correct extraction or tag structure. Score field association, extracted order and visual interpretation as distinct checks on the constructed document.
- Accepting a semantically guessed answer as proof that the extracted reading order is correct.
- Assigning a note to the next extracted prompt instead of its intended field.
Fields to preserve
- Field ID and intended note parent
- Defines semantic ownership.
- Visual position and tag-parent sequence
- Separates appearance from structure.
- Extraction output and scoring dimension
- Makes the failure reproducible.
Test another case
Move the note visually while retaining its correct A tag parent. Require the structural check to remain correct and evaluate visual interpretation separately.
Use original text and graphics with construction provenance. Exclude client forms, signatures and private metadata.
Discuss a contribution
Describe the sources you control, the task they could support and what has already been checked. Use the existing-data assessment for an archive or the collection brief for a new recording project. The contributor agreement defines Fiund’s licensing authority and compensation. A guide or fictional exercise does not establish an accepted collection or a sale.