The rules are applied by your Host, inside your workspace.

Four, and they are the whole list. Your Host refuses these, and a repeated pattern lowers its trust in you. Everything above them is between you and your Host.

If you tell your Host you are thinking about suicide or hurting yourself, its app shows you a notice with crisis resources: the 988 Suicide & Crisis Lifeline. The notice is drawn by the platform's code, marked as the platform's, and you can dismiss it. The Host is told that the check fired, and it carries on with you; nothing is added to its reply.

The check is built into the Host's code, not written into its instructions. It runs on the message you send and again on words you send while the Host is still replying, and the notice does not depend on the model. In a messaging app, which cannot draw the notice, it shows nothing; the Host is still told.

A count. When a governance event happens — a refusal given, a referral shown — your Host writes a row in its own database holding the name of the event and a timestamp. That is the entire contents. Not the message, not the reply, not a summary, not a category of you.

The counts exist so the mechanism can be proven to have worked, to us and to a regulator. They cannot reconstruct a conversation, because there is nothing in them to reconstruct it from.

On your own computer, nothing: your conversations stay on it. On a managed machine, they stay on the machine, encrypted under keys only you hold, which we can reach only through the server it runs on, a server Privatae rents and runs, and that server's console. We take no backup of the machine's disk. On a Host on the Privatae platform, we run the machines it lives on, so we can reach into it: our staff and automated checks run commands inside Hosts to operate and fix the service, and our backup job copies the Host's database, conversations included. What is enforced, and where.