TAR-set anonymization is the removal of personal data from documents fed to technology-assisted review. FRCP 26(b)(1) makes discovery proportional, which is the reason ranking is used at all. anonym.plus keeps the corpus on your own device.
When this applies
A ranking model reads every item in the collection, privileged ones included, before any lawyer does. Send that corpus to a hosted service and you have disclosed the whole collection. Proportionality under Rule 26(b)(1) is why you rank.
How anonym.plus handles it
- Point anonym.plus at the document set on your device.
- Local OCR reads any scanned items in the set.
- The tool flags names, contacts, and IDs across files.
- Use steady labels so relevance signals survive.
- Replace or mask each confirmed value.
- Save the clean set for the ranking workflow.
What you need to provide
- The document set (mixed PDF, DOCX, email, or scan).
- Replace with steady labels to preserve text signals.
- Optional batch run of up to 20 files at a time.
PII entity types detected
| Category | anonym.plus entity type | Example |
|---|---|---|
| Names | PERSON | custodian name → [PERSON_n] |
| Contact | EMAIL_ADDRESS | sender email → [EMAIL] |
| Dates | DATE_TIME | doc date → [DATE] |
| Identifiers | US_SSN | SSN → [SSN] |
| Location | LOCATION | address → [ADDRESS] |
| Account | US_BANK_NUMBER | account no. → [ACCOUNT] |
Compliance achieved
- FRCP 26(b)(1) proportionality is the ground for ranking instead of reading all.
- Nothing is uploaded, so the training corpus is never disclosed to a hosted model.
- Steady labels keep the text signal, so recall testing stays meaningful.
- Scanned items are read by local OCR before ranking begins.
Anonymize TAR document sets offline — see plans & start free →
Limitations & cautions
Anonymizing before machine ranking can shift how a model reads context. Steady labels keep most signals, but test recall on a control set first. Free-text clues that survive redaction still need a human pass on responsive items.
Frequently asked questions
Do courts accept technology-assisted review?
Yes. Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012) was the first opinion to approve predictive coding, and later rulings followed. Courts look at the validation protocol, not the brand of software.
Does anonymizing first hurt recall?
It can shift context. Steady labels preserve most of the signal, since one custodian keeps one token throughout. Validate on a control set and measure elusion before you rely on the ranking.
Why clear the data before the model trains?
A hosted model ingests the whole collection at once, privileged items included. Clearing first, on your own hardware, keeps that exposure at zero.