Page Number Extraction

Page number extraction is the reading of pagination from document pages — the "3", the "Page 3 of 17", the "iii", the "2-14" of sectioned manuals — and its conversion into structural knowledge the pipeline can act on. Pagination is the document's own integrity metadata: "Page 3 of 17" asserts sixteen siblings, their order, and this page's place among them — assertions a pipeline can verify, making page numbers one of the cheapest and most powerful validation signals multi-page processing has.

Extraction is a furniture-reading task with parsing attached: locating the pagination element (headers and footers, per those entries' repetition-and-position analysis — distinguished from dates, section numbers, and the other numerics that share the furniture zones by pattern and cross-page increment behavior), then parsing its scheme: plain sequence, X-of-Y, roman-numeraled front matter switching to arabic body, per-section numbering (2-1, 2-2…), and the Bates numbering of legal productions — a parallel numbering layer with its own semantics, prefix conventions, and evidentiary weight. Scheme inference across pages beats per-page reading: the sequence's pattern disambiguates individual OCR wobbles (the page reading "l7" between 16 and 18 is 17), and deviations from the inferred pattern are the findings.

The findings are what justify the entry: gaps (page 12 missing from the scanned pack — caught at intake, when the source can still supply it, rather than at decision time), disorder (the shuffled rescan detected and re-sequenced by numbering rather than guesswork), duplication (page 7 twice — a scanning double-feed or a substituted page, each worth different attention), and boundary evidence (numbering restarting at 1 announcing a new document to the bundle-splitting machinery). In legal and compliance productions the stakes formalize: completeness certifications lean on pagination audits, and Bates continuity is checked as a matter of course — the humble page number doing chain-of-custody work.