A representative pass on a 10 page agreement. The model below is what we remove by default. Custom rules cover everything else.
- Client and counterparty names, opposing counsel, principals, signatories.
- SIN, business numbers, account and reference numbers, postal codes, and street addresses.
- Phone and email of every named person, including external counsel.
- Internal file numbers, settlement amounts, and dates of birth.
Why this works for privilege, not in spite of it.
- Canadian residency, Canadian regulation.
- Built under PIPEDA and Quebec Law 25. The little we hold (auth, billing metadata) is on a roadmap to Canadian processors. Documentation provided on enterprise request.
- The draft never reaches us.
- Detection, redaction, tokenization, and encryption run in your browser. Anaris never receives the document binary. Optional AI augmentation sees only extracted text, and only if you opt in.
- Reversible by you, opaque to us.
- Tokenized redactions recover with a passphrase only the holder knows. The key derivation and cipher choices are public; the passphrase is yours. We could not return the original if we tried.
Three questions partners ask.
- Does the LLM still understand the document if half the names are gone?
- Yes. Names become stable placeholders, so the model keeps the relationships and the structure. Counsel A still acts for Plaintiff X. The model just does not know who they are.
- What if I need to recover the original for the partner meeting?
- Recovery is a passphrase, not a server round trip. The encrypted token map is embedded in the redacted PDF. Anyone with the passphrase can restore the original locally.
- Does it work on French pleadings?
- Yes. Detectors run the same model on English and French. We will not give that up.
Try it on the next draft.
Free while in Research Preview. No account needed to start.