Privacy and PII

NomOS screens text for possible personal data before the applicable governance rule decides whether to warn or withhold it.

What detection controls

The confidence threshold sets how certain the screening service must be. A higher value produces fewer, more certain matches; a lower value finds more possible personal data and may create more false positives.

The selected data types determine what is inspected. Person names, email addresses, phone numbers, and IBANs are active by default; you can also turn on credit card numbers and AHV numbers. The technical name beside each option is the identifier that also appears in evidence records and warnings.

A first and last name together (“Anna Meier”) always clears any usual threshold. A single word only counts as a name reliably when a cue stands shortly before it: a title such as “Dr.” or “Prof.”, “Herr”, “Frau”, “Name”, or the heading of a person list such as “Kundenliste”. Color words that double as surnames (“Braun”, “Grau”) and lists of things such as a “Preisliste” stay below the threshold without a person context.

A match is not yet the resulting action

The settings determine what counts as a PII match. On a match, NomOS withholds chat and agent output only in rooms of the data class that the rule names, C1 by default. In rooms of other classes, the output is allowed. For planned agent actions and imports, the same match can appear as a warning instead.

If the screening service is unavailable, protection remains fail-closed. The interface identifies the outage explicitly and does not claim that personal data was detected.

The privacy check detected a possible reference to a person. This is a detection category, not a confirmed identification. Rephrase the question without personal details, or ask the room owners which context is permitted. The withheld answer text is not displayed here.

Test changes before saving

The test field under Settings → Privacy uses the threshold and data type selection currently entered, even before they are saved. Matches show the affected text, the technical data type, and the score.

The test text is sent to the internal screening service for this check only. It is not stored and does not create an evidence record.