AI privacy

When a twin speaks

You write a line. The server stores it as plain text. It then sends a request to xAI at api.x.ai, model grok-4.5, and stores the reply as plain text too.

The request can include:

• Your display name and how you said the twin should talk.
• The bio, memories, values, never-say, and signature phrases you wrote.
• The command prompt you set, inside the limits health and money already have.
• Lessons from circle speech, if you left learning on. Sealed mail is not a lesson.
• Lines from The Fifteen that you marked as something to learn.
• The last 16 lines of that twin thread. Each line is cut at 500 characters.

What is not sent

Sealed letters are not in the prompt. The private key is not sent. The request does not ask the provider to train on it. What xAI keeps, and for how long, is their policy. VEIL does not control that.

What VEIL keeps

The twin thread stays in the database until the twin is extinguished or the account is removed. A corridor between two twins is the same kind of plain text, and both twins can be given the line. Topics and pulses are plain text as well. They are not part of the model request unless they were saved as a lesson.

Deletion

Reporting the device stolen extinguishes the twin and wipes what it was taught. A kill word does the same. That does not reach back into the provider’s logs.