document-redactor (England & Wales)

This is a fork of kipeum86/document-redactor — an excellent offline DOCX redaction tool originally built for Korean legal practice. This fork adds detection rules for the legal system of England & Wales, with a focus on clinical negligence proceedings and inquests.

All credit for the core architecture, security model, OOXML handling, and verification pipeline belongs to the original project. This fork only extends the detection rules.

What this fork adds

The upstream tool ships with detection rules tuned for Korean and US legal documents. This fork adds 20 UK-specific rules across four new files, without modifying any existing rules:

Identifiers (identifiers-uk.ts)

Rule	Example	Tier
National Insurance number	QQ 12 34 56 C	All
NHS number (Modulus 11 validated)	943 476 5919	All
UK domestic phone (all Ofcom formats)	07700 900123, 020 7946 0958, 0117 496 0123	All
UK postcode	SW1A 1AA, B2 4QA	Standard
GMC number (context-gated)	GMC No: 1234567	Standard
NMC PIN (context-gated)	NMC PIN: 12A3456B	Standard
UK driving licence	SMITH 861215 J99KA 12	Standard
Hospital number / MRN (context-gated)	MRN: RXH 123456	Standard
UK bank sort code (context-gated)	Sort Code: 12-34-56	Standard

Legal (legal-uk.ts)

Rule	Example	Tier
Court claim number (all divisions)	KB-2024-001234, county court refs	Standard
Coroner's reference (context-gated)	Inquest Ref: 2024-0123	Standard
UK legal context scanner	Claim No: ..., Inquest into the death of ...	Standard

Temporal (temporal-uk.ts)

Rule	Example	Tier
DD/MM/YYYY date (calendar-validated)	15/03/2024, 15.03.2024	Standard
DD/MM/YY short date	15/03/24	Paranoid

Entities (entities-uk.ts)

Rule	Example	Tier
NHS Trust / Health Board / ICB	Barts Health NHS Trust, Betsi Cadwaladr University Health Board	Standard
UK judicial titles + name	His Honour Judge Smith, Mrs Justice Andrews, HHJ Taylor	Standard
KC / QC + name	Sarah Jones KC	Standard
Medical professional titles + name	Consultant Smith, Staff Nurse Patel	Paranoid
Medical record context labels	Patient:, D.O.B:, GP:, Ward:	Standard
Inquest context	Touching the death of, Deceased:, The late	Standard

What is deliberately not detected

Neutral citations ([2024] EWHC 123 (KB)), law report citations ([2024] 1 WLR 123), statute references (s.11 Limitation Act 1980), and CPR references are not flagged. These are public legal knowledge — they don't identify any person, case, or place.

What the upstream provides (unchanged)

Everything else comes from the original project:

• Zero-network architecture — CSP default-src 'none', ESLint network bans, build-time ship gate
• Single HTML file — download, double-click, redact
• OOXML deep traversal — body, headers, footers, footnotes, endnotes, comments, metadata, relationship files
• Round-trip verification — the output DOCX is re-parsed and checked before download
• Metadata stripping — scrubs author, company, tracked changes, comments, custom properties
• Field and hyperlink flattening — catches hidden URLs in OOXML instruction text
• Manual additions — type any string to add it as a redaction target
• 1,700+ automated tests with 90% coverage thresholds

See the upstream README and USAGE.md for full documentation.

Quick start

Use the built tool

1. Go to Releases and download document-redactor.html
2. Double-click to open in your browser
3. Drop a .docx file
4. Review candidates, add any the tool missed
5. Click Apply and verify
6. Download the .redacted.docx

Build from source

git clone https://github.com/banterny/document-redactor.git
cd document-redactor
bun install
bun run test
bun run build
open dist/document-redactor.html

Syncing with upstream

This fork tracks the upstream main branch. To pull in future improvements:

git fetch upstream
git merge upstream/main

The UK rules live in separate files (*-uk.ts), so merge conflicts should be rare.

Licence

Apache 2.0 — same as the upstream project.

Acknowledgements

This fork exists because kipeum86 built something genuinely excellent. The security architecture (defence-in-depth with three enforcement layers), the round-trip verification pipeline, and the single-file distribution model are all outstanding engineering decisions. This fork just teaches it to recognise UK postcodes and NHS numbers.