Voice Control
Voice turn safety, degraded mode, barge-in handling, and confirmation rules
Voice Control
Phase 11.3 adds the canonical voice-control runtime for safe voice escalation and control handling.
Voice is a control-intent surface, not a shortcut around governance. The runtime preserves the same canonical route, escalation, endpoint-trust, operator-control, and MAO truth already used by chat, Projects, and other in-app surfaces.
What Voice Control Does
- Tracks canonical turn state for a voice session
- Evaluates whether a turn is ready for handoff using combined signals
- Defers risky actions to clarification, text confirmation, or dual-channel confirmation when required
- Stops assistant output immediately on barge-in and records interruption timing
- Keeps degraded-mode status visible across operator surfaces
End-of-Turn Safety
Voice actions do not execute on silence alone.
The runtime evaluates all of the following together:
- semantic completion
- silence-window timing
- explicit handoff support
If those signals do not line up, the system continues listening, asks for clarification, or blocks the action instead of guessing.
Confirmation Rules
Voice can request an action, but it cannot bypass canonical confirmation policy.
- Low-confidence risky actions can require explicit text confirmation
- Destructive, critical, or
T3actions require dual-channel confirmation from the active Principal session - Voice alone cannot authorize those actions
When confirmation is required, the runtime preserves that posture in canonical session state so MAO and other in-app surfaces can show the same pending confirmation status.
Barge-In and Continuation
If you interrupt assistant speech:
- assistant output stops immediately
- interruption timing is recorded
- the session moves into continuation-required posture
The assistant does not automatically resume after interruption. Continuation must be explicit.
Degraded Mode
The runtime can enter degraded mode when voice handling is unsafe for risky controls.
Common reasons:
- low ASR confidence
- low intent confidence
- handoff instability
- transport degradation
- interruption-recovery posture after barge-in
While degraded mode is active:
- risky controls shift to text-first behavior
- operator surfaces keep showing current voice posture
- a single improved turn is not enough to clear degraded mode
This is a safety mechanism, not a cosmetic warning.
Privacy and Evidence
Phase 11.3 uses a metadata-first posture for voice control events.
By default, the runtime keeps:
- transcript hashes
- confidence and handoff signals
- decision refs
- route and escalation refs
- witness-linked evidence
Raw audio is not retained by default.
Current Scope
Phase 11.3 ships the canonical runtime and web/API composition seam. Phase 11.5 extends that runtime into the mobile operations surface as a projection-only follow-up view.
Included now:
- canonical voice turn evaluation
- interruption and continuation handling
- degraded-mode projection
- text-first and dual-channel confirmation posture
- mobile operations visibility for continuation-required and degraded-mode follow-up posture
Not included yet:
- connector-native voice call UX
- wake-word or ASR model tuning
- voice-only authorization for dangerous actions
The mobile surface does not evaluate turns locally. It reads the same canonical session projection already used by MAO and other web surfaces, so interruption and confirmation posture stays consistent across surfaces.