One catalogue. Every class is listed with how it is found, how it is validated, and what its token looks like. Coverage is measured per class on your own data before go-live.
All layers run under fixed model versions and thresholds, so the same input always produces the same output. No generative model rewrites your text; every replacement is rendered from the original bytes.
Deterministic. A number has to pass its own validation before it counts; a labelled value that fails, for example after OCR damage, is still masked at reduced confidence.
Purpose-built models, each with one job: names, organisations, addresses and places. Trained on our own data, multilingual.
Reads scanned pages, photos and image attachments on your servers. Signatures, identity documents and faces are blocked outright; printed text goes through the same rules and models as any document.
A second model re-checks every candidate in its context, including the row and header of a table cell, before anything is replaced.
Resolves the class when layers disagree, then applies the tenant policy: replace and restore, stop, or leave readable.
Who the text is about.
A natural person named in the text, with an attached title when present.
The same person keeps the same token for the whole case.
Companies and institutions named in the text.
Measured as its own slice. Tenants decide whether public institutions are masked.
Turkish Republic identity number.
A labelled number that fails its checksum, for example after OCR damage, is still masked at reduced confidence.
Turkish and foreign passport formats when a passport label is nearby.
Birth dates only. Other dates stay readable so the model can reason with them.
Turkish registration plates.
How to reach or find them.
Mobile and landline numbers, including subscriber numbers in traffic records.
Any address, including Turkish characters in the local part.
A deliverable address: street, number, district, with or without the city.
A place named on its own: a city, district, country or administrative chain. Never a postal address.
Tenants that want cities readable switch this class off.
Latitude and longitude pairs in decimal or degree notation.
Only links that identify a person, such as a profile page. Other links stay readable.
IPv4 and IPv6 addresses.
Numbers that must pass their own checksum before they count.
International bank account numbers, Turkish and foreign, with or without separators, in any case.
Debit and credit card numbers.
Default treatment is stop: the request is refused rather than tokenised, unless the tenant configures otherwise.
Turkish tax identification numbers of companies and sole traders.
Domestic account numbers when an account label is nearby.
Subscriber, customer, contract, policy and member numbers, 6 to 20 digits, gated on the label word.
Subscriber and handset identities as they appear in traffic records.
Subscriber identity on the SIM.
SIM card serial number.
Device identity of a handset.
Foreign identity numbers, secrets, network and sector identifiers. Each has its own validator, reference tests and negatives on ordinary text.
German tax ID, Spanish NIF and NIE, US SSN, Canadian SIN and the other formats that appear in your documents. Each with its own checksum and reference tests.
Turkish and foreign licence numbers in identity-document workflows.
Hardware addresses in telecom and network logs.
API keys, access tokens and private keys, built on established secret-detection rules. Covers the coding and chat relay.
Routing numbers, sort codes and similar domestic formats.
Patient and insurance numbers.
Wallet and exchange addresses across the major chains.
Found on scanned pages, photos and image attachments. These are never tokenised: the request is blocked before anything reaches a model.
A signature on a scanned page, photo or image attachment.
Passports, national ID cards, driving licences and other identity documents, scanned or photographed.
A recognisable face in a photo or on a scanned page.
Your own identifiers, as a first-class class.
Case numbers, policy numbers, internal IDs, registry-backed values. Defined per tenant or globally, activated at runtime, same tokens and audit as every other class.
Optional and complementary. Masks a stated disclosure about an identifiable person; generic statements pass unchanged. It only adds protection; it never removes or re-types what the core found.
The semantic layer. Two purpose-trained models work together: one screens a passage for a sensitive category, the other locates the stated disclosure in the original text. Salus validates the offsets, merges the spans with the core result and renders one protected text. The layer only adds protection. It never removes, re-types or re-bounds what the core found, and it never renumbers a token. Customer policy selects the enabled categories and whether an incomplete semantic pass holds the request or lets the core result through.
A stated illness, treatment, pregnancy or disability of a particular person.
A stated belief or practice of a particular person.
A stated fact about a particular person.
A stated age or gender of a particular person, beyond a bare date of birth.
A stated nationality of a particular person.
Marital status, children, dependants of a particular person.
Where a particular person works and what they do.
Income, debt, arrears of a particular person.
Where a particular person is or lives, said in prose rather than as an address.
All examples are synthetic.