Placeholders explained

A placeholder replaces a real value in the text the AI receives. You choose the style.

Style 1: stable tokens (default)

Jean Dupont  →  <PERSON_558644>
jean@acme.fr →  <EMAIL_ADDRESS_094087>

• The same value always gets the same token, in every message, file and session. The AI can follow who is who, and caching keeps working.
• Tokens are computed with a secret key kept on your computer (~/.local/state/pii-airlock/token.key). Without it a token tells nobody anything about the value behind it.
• Tokens change over time. By default a person gets a new token each day, so a provider cannot follow one person across all your conversations.

token_rotation: day     # day (default), week, or never

Use never if you clean one big dataset over several runs and need identical tokens throughout.

To change all tokens immediately, delete token.key.

Style 2: realistic fakes

token_style: surrogate

Jean Dupont         →  Jasper Iverson
jean.dupont@acme.fr →  jasper.iverson@example.com

The text reads naturally, which some models handle better than <TOKENS>. Phone numbers, cards and IBANs keep a valid shape. Other types, such as secrets and places, still use tokens.

Trade-off: if the AI invents a name that happens to equal a fake, that name is also swapped back.

Style 3: numbered

token_style: sequential

<PERSON_1>, <PERSON_2> …

Numbered in reading order, separately for each request or file. This was the default in early versions.

Where do I put these settings?

In ~/.config/pii-airlock/config.yaml for yourself, or .pii-airlock.yaml in a project. See Settings.