Since v1 the tutor has been non-streaming — you'd wait for the full reply to appear at once. As of this week, replies stream token by token.
The engineering challenge was keeping the leak-guard synchronous with a streaming reply. We solved it by streaming to the client and running the leak-guard in parallel on the server. If the guard flags a leak mid-stream, we swap the visible reply for the refusal template with a subtle fade. It happens in roughly one in two hundred replies.
The UX is better across the board. Sessions feel alive. The composer feels present. Even users who never noticed the wait notice the presence.
Filed by Engineering on Apr 8, 2026. Under Engineering.
the journal is where we think in public — reply-friendly