Office Document Sanitisation

Overview

When OpenRegister anonymises an Office document (.docx or .odt), the entity walker can only reach the visible text runs. Office documents also carry personally-identifying information (PII) in non-text structures that the walker is structurally blind to — reviewer comments, tracked-change author attributes, document metadata, person-identifying field codes, custom XML data bindings, and hyperlink URLs. These regularly leak names and case numbers into published Woo/dossier documents.

The document sanitiser runs ahead of the anonymisation walker. It produces a cleaned derivative of the input by performing XML-level surgery on the ZIP container; the cleaned file is then passed to the walker for entity detection and replacement. The original file is never modified.

What it strips

Category	DOCX	ODT
Comments	word/comments*.xml + people.xml + inline commentRange*/commentReference markers	office:annotation / office:annotation-end
Tracked changes	accept <w:ins> (keep text), drop <w:del> (remove text), strip w:rsid* revision attributes	accept text:change-start..end, drop text:change, remove the text:tracked-changes container
Document metadata	docProps/core.xml, app.xml, custom.xml (Author, Last-Modified-By, Title, Subject, Keywords, Description, Category, Content-Status, Company, Manager, string-typed custom props)	meta.xml (creator, initial-creator, title, subject, keyword, description, string-typed user-defined)
Person field codes	AUTHOR, USERNAME, USERINITIALS, LASTSAVEDBY (simple <w:fldSimple> + complex <w:fldChar> forms)	text:author-name, text:author-initials, text:initial-creator
Custom XML	all customXml/item*.xml parts; data-bound <w:sdt> content controls are unwrapped (visible text preserved)	—
Hyperlinks	<w:hyperlink> flattened to plain text; URL + relationship dropped	text:a flattened to plain text

Timestamps (dcterms:created/modified) and non-person field codes (DATE, PAGE, etc.) are preserved. Removed parts are reconciled out of [Content_Types].xml and _rels/*.rels so the output opens cleanly in Microsoft Word and LibreOffice (no "found unreadable content" recovery).

Metadata sentinel (design D5)

Scrubbed metadata fields are replaced with the sentinel string DocuDesk Anonymisation rather than deleted. Keeping the element with a recognisable value (a) signals in-file that the document was processed, and (b) defends against Word's "fill missing metadata on save" behaviour, which would otherwise re-populate <dc:creator> with the current user on the next save — re-leaking PII. The sentinel is a single constant and can be changed to a generic value (e.g. Anonymous) without touching the surgery logic.

Hyperlink flattening (design D7)

Hyperlinks are flattened to their visible text; the URL (which often carries PII such as mailto:p.jansen@… or query strings with case numbers) and its relationship entry are dropped. The visible link text remains and is then walker-anonymised in the entity pass.

Audit report

Each sanitisation produces a SanitizationReport with per-category counts (comments removed, tracked changes accepted/dropped, revision attributes stripped, hyperlinks flattened, metadata fields scrubbed, custom XML parts dropped, field codes stripped, sentinel applied). The report is PII-free (counts only — never document content) and is retained on DocumentProcessingHandler::getLastSanitizationReport(). Logging follows ADR-005: only file ID, MIME type, strategy class and counts are logged.

Encrypted documents

Password-protected .docx/.odt cannot be sanitised without the password. The sanitiser raises a typed SanitizationException (reason encrypted); the anonymisation flow surfaces this as a caller-correctable "Cannot anonymise an encrypted document" error.

Scope

• One-way only — the original file is the source of truth for the un-sanitised content; there is no reverse operation.
• v1 is always-on, all categories (no per-tenant or per-category toggles).
• Internal-document anchor hyperlinks (#bookmark1) are flattened along with external ones in v1.