A realtime speech-to-speech model can be verified — REALTIME_AUDIO is probed by opening a session (voice realism VR1, D9)

Shipped

October 26, 2026 at 8:25 PM UTC

Author

Kamo

Commit

1438893

REALTIME_AUDIO was ATTESTED, and only platform providers can be attested, so no org's realtime model could ever be VERIFIED, and an SI is routed to VERIFIED models only: the speech-to-speech pilot the owner ordered for crm (D9, 2026-10-26) could not get a ticket. New probe kind REALTIME_SESSION, appended last (no existing value moves); AIService's ProbeRunner implements it.