Context preservation

One sentence, three treatments

Original. Elif called about Elif’s account.

Redacted. ▮▮▮ called about ▮▮▮ account. Two markers, zero information: the model cannot tell whether they are the same person, and nothing can restore them.

Tokenised. PERSON_rqmzal called about PERSON_rqmzal’s account. The repeated token preserves continuity: one person, referenced twice, class known.

What the model actually needs

Most enterprise AI tasks are not about the sensitive value itself. They are about the structure around it. Summarise the complaint. Draft the reply. Decide which team should own the ticket. For those tasks the model needs to know that the caller and the account holder are the same person, that the account is an account, and that the third mention of the customer is still the same customer. It rarely needs the digits of the phone number.

Typed, stable tokens carry exactly that structure, through a long document, a multi-turn conversation or a batch of records.

How Salus keeps one entity one token

Inside a context the engine clusters spans by class and normalised value. Normalisation folds case, whitespace and the Turkish İ/ı distinction, so a name typed three ways is one cluster. A bare surname joins the one full name in the text that contains it; a surname and given name split across two captioned form fields are joined; the name fields of a machine-readable passport zone are joined to the printed name. Every member of a cluster gets the same token.

The window the engine cuts text into for detection overlaps its neighbours, so an entity that straddles a boundary is not split into two.

An honest boundary

Context preservation is about structure, not magic. The model still loses the value itself, so operations that need the real characters degrade, as Tokenisation explains. The claim worth making is precise: this preserves more useful structure than deletion or undifferentiated masking, and for the reasoning tasks enterprises actually run, that structure is usually what quality depends on.