Method / Note 2
How Every Quotation Is Verified and Page-Anchored
The verification standard that keeps a quotation tied to the exact as-published source, page, and edition.
At a glance
- The source edition is registered before a quotation is extracted.
- Page boundaries are preserved throughout text extraction and retrieval.
- Quoted text is matched character for character after only controlled normalization.
- The citation identifies the printed page and the PDF page of the as-published edition, not an untraceable excerpt.
- A human reviewer confirms the quotation, page, context, and evidentiary use.
Why quotation verification matters
A quotation carries a stronger claim than a paraphrase. It tells the reader that the source used those words. In a litigation or regulatory research product, a misplaced word, missing qualifier, wrong page, or quotation drawn from the wrong edition can materially change the meaning.
For that reason, RecordStrata treats quotation verification as a provenance problem, not a copy-editing step. The process must answer four questions: Which document? Which edition? Which page? Which exact characters?
Quotation marks are a representation about the source. They are used only when that representation can be verified.
1. Register the exact source edition
Before extraction begins, the source PDF is registered with its report number, filename, page count, retrieval information, and cryptographic hash. The sample Deep-Reads identify the as-published investigation report and its MD5 hash. That hash functions as an edition fingerprint: it distinguishes the document used from a later replacement or revision with the same visible title.
The source file is retained unchanged. Any extracted text, corrected OCR, page image, or structured field is stored separately. This permits the extraction to be improved without altering the historical evidence object from which the quotation was taken.
2. Preserve page boundaries during extraction
Text extraction must retain the association between every passage and its source page. A single flat text file is not sufficient because it can lose page divisions, headers, footers, table position, and the distinction between body text and captions.
RecordStrata therefore treats the document as a sequence of page objects. Each page can contain native text, layout coordinates, tables, and visual material. Where a report page is image-based or contains an annotated diagram, page-image review is used rather than relying on OCR alone.
3. Separate quotations from paraphrases
| Presentation | Required treatment |
|---|---|
| Verbatim quotation | Must match the registered source text, be enclosed in quotation marks, and carry a source-page reference. |
| Close paraphrase | Must preserve the meaning, remain attributed to the source, and avoid wording that could be mistaken for a quotation. |
| RecordStrata synthesis | May combine supported facts from more than one passage, but must be labelled as synthesis and must not be placed in quotation marks. |
| Analytical label | May organize the record – for example, "management-knew record" or "dual citation" – but must be identified as RecordStrata terminology. |
| Unverified text | Must not be published as a quotation. It may be held for review or rewritten as a carefully supported paraphrase. |
4. Character-level matching
The quotation-checking process compares the proposed quotation with the text extracted from the registered page. Only controlled normalization is permitted. Examples include joining a word broken by a line-ending hyphen or normalizing a typographic apostrophe where the visual character is the same. Normalization must not delete qualifying words, repair grammar, modernize capitalization, or silently substitute a preferred phrase.
A successful match records the document identifier, page identifier, matched text, location within the page representation, and verification result. If more than one page contains the same phrase, the citation is resolved using context and the section in which the passage appears.
What character-level verification is designed to catch
A missing word such as "not" or "could not determine."
A changed number, unit, date, or cited standard.
A quotation copied from a summary rather than the official report.
A sentence assembled from non-contiguous passages.
A quote attributed to one page when it appears on another.
Text drawn from a later or different edition of the source.
5. Page anchoring
A page anchor tells the reader where the supporting language appears in the exact source edition. RecordStrata cites the page number printed on the source page, followed by the PDF page in brackets: for example, [MSHA Report of Investigation, p. 4 (PDF p. 6)]. The printed number is the one counsel will cite; the PDF number is the one a viewer opens. Where a page carries no printed number, such as a cover, the citation says so and gives the PDF page.
The page anchor should remain stable even if the quotation is later reused in a report, an Ask the Record answer, or a linked citation viewer. The user should be able to open the cited page and inspect the surrounding text, table, photograph, or diagram rather than receiving an isolated excerpt.
6. Context verification
An exact string match is necessary but not sufficient. The reviewer must confirm that the quotation is used in the same sense as the source. A passage may appear in an interview account, a discussion section, a root-cause finding, or enforcement language; those locations carry different evidentiary weight.
7. Human verification
The sample Deep-Reads state that quotations were machine-verified against the source PDF character for character and page-anchored automatically, after which the extraction record was verified line by line by a human reader. That human review is the release gate.
The reviewer checks the visible PDF page, confirms the quoted wording and punctuation, inspects the surrounding paragraph, verifies the page reference, and confirms that the quotation supports the proposition for which it is used. The reviewer also checks whether the source attributes the statement to an interviewee, an investigator, the operator, or another participant.
8. What happens when verification fails
| Failure | Required response | Publication result |
|---|---|---|
| No exact match | Recheck the edition, page, extraction, hyphenation, and proposed quotation. | Do not publish as a quotation until resolved. |
| Image-only text | Inspect the page image and, where needed, create reviewed OCR tied to that page. | Publish only after visual confirmation. |
| Ambiguous duplicate phrase | Use surrounding text and section context to resolve the correct occurrence. | Cite the confirmed page and context. |
| Source differs from draft | Correct the draft; do not "correct" the source. | The source wording governs. |
| Material conflict between passages | Preserve both passages and label the conflict. | Do not select one merely for narrative simplicity. |
| Unable to verify | Remove quotation marks or omit the passage. | Use no unsupported verbatim claim. |
9. Quotations in Ask the Record
Ask the Record may quote a source passage, paraphrase it, or synthesize multiple supported passages. The answer must make that form visible. A direct quotation should be short, exact, and linked to its page. A synthesis should identify the documents and pages relied upon and should not be presented as though the source had written the answer in that form.
The citation viewer should open the page itself, not only the extracted text. Where the answer depends on a photograph, diagram, or table, the cited page should display that visual evidence. The user must be able to verify the response against the original.
10. Revisions, corrections, and later editions
If an agency later publishes a revised report, RecordStrata should register it as a new edition. Existing quotations remain tied to the edition from which they were taken. A later editorial revision may update the analysis, but it should not retroactively erase the source history.
If a quotation or page anchor is found to be wrong, the correction should identify the affected publication, the nature of the error, the correction date, and whether the substantive conclusion changed. The corrected extraction and the prior published version should remain traceable under the corrections policy.
Publication standard
No quotation without an identified source edition.
No quotation without an exact or visually confirmed match.
No quotation without a page anchor.
No quotation without context review.
No quotation that changes the source to improve grammar or clarity.
No machine-verified quotation released without human review.
A citation is not decoration. It is the route by which the reader can test the claim.
Source notes
The examples in this Method Note are drawn from the following records. Page references are to the cited as-published or Deep-Read edition.
1. "A Berm That Management Chose Not to Build," MSHA Fatality Deep-Read Series No. 1, Nevada Gold Mines, LLC, dated September 7, 2026, especially the dataset note on p. 4.
2. "Eleven Minutes: A Coordination Failure, Cited Twice," MSHA Fatality Deep-Read Series No. 2, Peabody Bear Run Mining LLC, dated September 7, 2026, especially the factor table on p. 3 and dataset note on p. 5.
3. MSHA Report of Investigation, Wendling Quarries Inc., Portable Crushing Department 415, fatal slip/fall accident of January 8, 2020, especially PDF pp. 10-11.