Settings (optional)
Sends the same audio to Deepgram and routes extraction to the in-tenant 12B instead of Anthropic. It isolates the reasoner; it is not a sovereignty configuration, because the audio still leaves the tenant. The endpoint has been live since the vLLM-FP8 cutover, so this no longer falls back to cloud in normal operation — a served by badge below reports which reasoner actually answered.
Replaces names/addresses/PHI with per-incident tokens (PERSON_1, LOC_1) before the payload leaves the tenant, then maps them back before anything reaches your screen. "Server default" sends nothing and uses the backend's ARES_ENABLE_PAYLOAD_DEID — pick On/Off to A/B it per transmission. No effect on the benchmark arm (the in-tenant reasoner is inside our own tenant) or on Edge (nothing egresses).
Known context for this incident —
dispatch, building, comms plan. "none" is the production default.
Idle. Hold the button and speak — radio traffic is extracted; a spoken question is answered.