When a PDF is uploaded to the Workbench, the platform parses the document's internal object structure — extracting the page tree, existing bookmark hierarchy (PDF outline dictionary), annotation objects, and embedded metadata. This parsing happens server-side in a sandboxed processing container and typically completes within 4–12 seconds for documents up to 500 pages. The result is a structured internal representation of the document that powers every downstream tool in the Workbench without requiring the user to wait for full page rendering.

The AI Bookmark Advisor sends the document's extracted text structure and existing outline to the DNXT regulatory language model, which has been fine-tuned on eCTD document hierarchies, ICH M4 module structures, and agency reviewer navigation patterns drawn from FDA guidance on electronic submissions. The model compares the current bookmark set against expected navigation points for the document type — clinical study report, SmPC, investigator brochure, etc. — and generates a ranked list of bookmark suggestions with confidence scores and proposed anchor page numbers. Suggestions flagged above 85% confidence are pre-selected; lower-confidence items are presented for manual review.

The blue-text detection engine performs a pixel-level and text-layer dual scan of every page. The text-layer pass uses regular expression patterns trained on regulatory cross-reference formats — section references, appendix citations, module path notation, and external URL patterns. The pixel-layer pass independently identifies rendered blue-coloured text regions using colour space thresholding, catching hyperlink candidates in scanned documents or image-heavy PDFs where the text layer is absent or unreliable. The two pass results are merged and de-duplicated, producing a comprehensive map of all potential hyperlink locations with page number and bounding-box coordinates. This dual-pass approach is why DNXT achieves 100% surface coverage where single-pass text extraction routinely misses 8–15% of link candidates.

Each detected hyperlink candidate is resolved against the submission's file index. For internal cross-document links, the platform checks whether the target file exists in the dossier, whether the target page number falls within the document's page count, and whether the target bookmark anchor is present in the destination document. For external URLs, the platform performs a HEAD request to check reachability (where policy permits) and flags any URLs pointing to non-persistent resources, which are inappropriate in a regulatory submission. Each link receives a status: Resolved, Broken Target, Missing Anchor, or External — surfaced visually in the viewer with colour coding so the publishing specialist sees the problem directly on the page where it appears.

The publishing specialist reviews each flagged link directly in the browser viewer — clicking Accept confirms the link annotation will be written to the output PDF, clicking Reject removes it, and Defer holds it for escalation. Every decision is written to the Workbench session audit log with a UTC timestamp, user identity, document name, page number, link text, and the decision made. This event stream is tamper-evident and exportable as a signed PDF audit report. For teams operating under 21 CFR Part 11, this provides the complete electronic record of who reviewed what and when — a record that desktop Acrobat workflows are constitutionally incapable of producing.

When a scanned document is detected — identified by the absence of a text layer covering more than 60% of pages — the Workbench automatically routes the document through the scan enhancement pipeline. This applies adaptive deskewing, noise reduction, and contrast normalisation before running optical character recognition at the page level. The resulting searchable text layer is embedded into the PDF without altering the visual rendering of the original scan, making the document compliant with FDA's requirement for searchable text in electronic submissions. The original scan pixel data is preserved in a separate archival layer so the enhancement is non-destructive and the output is auditable as a derivative document.

When the publishing specialist completes the review, the Workbench writes all accepted bookmarks, confirmed link annotations, and embedded text layers into a new PDF/A-compliant output file. The export also packages a Workbench Session Report — a structured JSON and human-readable PDF document listing every change made, every link reviewed, and the full audit trail — which travels with the submission package for QA sign-off. The output PDF is validated against the platform's eCTD technical specification checker before download, catching issues like invalid PDF version flags or non-conforming bookmark structures before the document ever reaches the submission authoring tool.