A sentence is missing from the comparison. Is it missing from the page?
The manuscript says “The workshop does not open on Mondays” while the proof extract says “The workshop opens on Mondays”. A sentence about collection is also absent from the extract. These are different flags: changed wording in the supplied texts and a passage not yet located. Before editing the layout, you need to establish what is actually in the PDF.
This guide proposes trialling a text assistant to prepare a comparison checklist with quotations and specific questions. It includes a fictional case and a manual reference for checking the response. We have not run the pilot or measured results. The assistant is not inspecting a PDF here, recomposing a book or approving a file for print.
Keep three checks separate
Adobe documents that InDesign preflight can flag overset text among other problems. It also allows configuration of the content and page ranges checked. Review the profile and scope for your document: comparing text does not replace that check.
Separately, the Poppler pdftotext manual distinguishes preserving physical text layout from producing text in reading order. Adobe explains that a tagged PDF’s reading order can change without altering its appearance. In this exercise, a block appearing elsewhere in an extract is therefore not treated as having disappeared.
The third check is editorial: comparing wording, numbers, notes and passages against the relevant manuscript version. AI would be trialled as help with organising differences and questions; the responsible person must locate passages in the proof and inspect the layout document. No integration with InDesign, Acrobat or Poppler is presented as already available.
Prepare a small package with a version and scope
Work on an agreed fragment including its associated notes or boxes. Identify the manuscript, PDF proof, date or version of each file and pages reviewed. If you provide only an extract, state that explicitly: do not let a response describe the whole book as checked.
Keep a literal copy first. If you prepare another copy to compare line breaks or split words, record your changes and keep the unnormalised text to verify every flag. Do not automatically remove negations, numbers, note references or meaningful punctuation as part of cleaning the text.
- Manuscript approved for this comparison: identified version and explicit scope, without substituting a later draft.
- Proof and extraction: version name, included pages and method used; a page with no extracted text is not treated as an empty page.
- Locators: section, opening words and page only when supplied by the source. Exercise identifiers are comparison labels, not paragraph numbers discovered by AI.
- Flag status: textual difference, not located in the extract or awaiting a check in the proof. Editorial approval belongs in a separate field.
Fictional case: a negation, an absent sentence and a displaced note
All texts, versions and locators below were invented by the editor. They are not client documents or real output from an extractor or assistant. O1 represents a complete fragment of a fictional leaflet manuscript. E1 represents the supplied extract of pages 12–13 of proof P1. No real PDF was created or inspected for this example.
O1 is ordered A, B, C and note N. E1 is supplied in the order A, N and C; B does not appear in that extract. Letters are exercise labels, and pages are supplied only where stated:
| Passage | Manuscript O1 | Extract E1 from P1 |
|---|---|---|
| A | The workshop does not open on Mondays. | Page 12: The workshop opens on Mondays. |
| B | Collection is agreed in writing. | Does not appear in the supplied extract. |
| C | Confirmed jobs are kept in a shared list. | Page 13: Confirmed jobs are kept in a shared list. |
| N · fragment note | Opening hours subject to confirmation. | Page 12, before C in the extracted sequence: Opening hours subject to confirmation. |
The manual reference records what is known and what needs inspection
A contains a literal difference: “does not open” in O1 versus “opens” in E1. The negation must not be dismissed as formatting. A useful flag asks for comparison against P1 and the layout file before deciding a correction. E1 alone does not establish what appears on the page.
Record B as “not located in E1”, quote O1 and request a search in P1. Do not claim that the typesetter deleted it or invent a page where it is missing. Absence from an extract does not yet establish whether the cause is an overset frame, a box, incomplete extraction or something else.
C matches literally in the case inputs. N also matches in wording, although it appears before C in E1. Do not duplicate N, rewrite it or declare it incorrectly positioned in the layout based only on that sequence. Its placement and relationship with its reference still need checking in the proof.
The working question becomes: “In P1, check the negation in A against O1, locate sentence B and check how note N relates to its reference. E1 does not establish that the composition is approved or B has been lost.” This is a handwritten internal reference, not a message sent to anyone.
An instruction to trial without rewriting the author
Paste the fictional O1 and E1 inputs, including their scope, into an authorised text assistant. Anthropic recommends allowing uncertainty and requesting citations to make a response auditable. We apply those principles to this trial design without guaranteeing the assistant will follow them.
“Prepare a comparison checklist using only O1 and E1. For each passage A, B, C and N, return an exact manuscript quotation, exact extract quotation when present, supplied version and locator, difference and pending check. Distinguish different wording from wording not located in the extract. Preserve negations, numbers and punctuation. Do not confuse extraction order with visible page order. Do not invent quotations, pages, causes or absent text; do not rewrite, summarise or improve the manuscript. Do not approve the layout or declare the PDF corrected. Return a draft for human comparison and state explicitly what you cannot verify.”
Check the response against the manual reference above. If it proposes deleting N to align the sequences, treats B as lost or adds the negation directly to a file, it has exceeded the task. Correct the draft and record the failure: an orderly list does not establish that the comparison is correct.
A human check arrives: update the finding, not the manuscript
The second phase is also fictional. Add record H1, linked only to proof P1: “I located Collection is agreed in writing. in a box on page 13 of P1. It does not appear in E1. I have not checked A or the reference for N.” This is an invented observation for the exercise, not an inspection we performed.
Ask for an updated checklist using H1. The manual reference changes B to “located in P1 according to H1; absent from extract E1”. It should no longer request reinserting B as if its loss were established. Retain E1 and H1 and, in a real job, record who supplied the check.
A still needs comparison and N still needs checking against its reference: H1 resolves neither. It does not approve all of P1. If a new proof P2 arrives, create a comparison for that version; do not automatically carry an observation about P1 over to P2.
Measure the complete comparison before automating
Start with the invented case and prepare your own reference before asking for a response. Also trial a variant without E1: the assistant should request the input and state it cannot compare, without reconstructing it from the earlier example. These are proposed tests, not observed results.
Compare manual reading supported by a table with the complete assisted process: preparing texts, reading differences, locating passages and correcting flags. A conventional text comparison tool may suffice for literal differences; the reason to trial AI is whether it helps describe pending checks and relate notes without inventing them.
- Record missed material differences, invented quotations or pages, and flags that confuse extraction with loss from the layout.
- Check that H1 updates only B and leaves A and N open. Count unsupported closures and questions that were already resolved.
- Record total preparation, comparison and correction time for both procedures. Do not claim savings or accuracy before measuring them.
- If a response hides a negation, alters the manuscript or presents an incomplete proof as approved, discard it and review the procedure with the editorial lead.
When to explore a narrowly scoped editorial integration
The manual trial copies fictional fragments and returns a list without changing files. A connected automation would need to identify versions, preserve manuscripts and extracts, link every observation to its proof and retain human approval. Actual exports and permissions in the editorial workflow would need checking before proposing compatibility.
Adaris Tech can assess custom software, CRM or automation case by case: for example, an internal panel gathering the original passage, extract, flag and responsible person’s check. This is a proposal to define, not an available proprietary editorial product or a book-approval tool. A first delivery should define a document type, access, retention, recovery and permitted actions.
Use fictional data only in this exercise. Before handling real manuscripts, check permission, usage rights and suitable confidentiality and processing conditions; do not paste unpublished manuscripts or personal data without authorisation. To discuss your case, describe the workflow without attaching client content. Commercial support is in Spanish.
Related services
Sources
Do proof versions leave observations in different files?
Tell us how you compare versions and who approves corrections. We can assess a narrowly scoped internal workflow; commercial support is in Spanish.