Voice input and output posture
What you'll find here
The VoiceSurface contract — how voice input and output is bounded
as a governed operator surface, not an always-on assistant layer.
No always-on assistant behavior.
Voice support is a governed extension point. Always-on or raw-audio-retaining behavior is deferred.
What it records
Surface id
Direction
input · output · bidirectional.
Support level
supported · experimental · deferred.
Consent requirement
Policy context preservation
Evidence link preservation
Receipt requirement
Transcript redaction
Media metadata redaction
Raw audio payload persistence
Documentation reference
Decision rules
Allowed (supported)
- Explicit operator consent
- Transcript redaction
- Media metadata redaction
- Policy context
- Evidence links
- Receipts
Blocked
- Persists raw audio payloads
- Skips operator consent
- Omits transcript or media metadata redaction
- Loses policy or evidence links
- Skips receipts
Redaction boundary
Voice records preserve ids and summaries — never raw audio or private transcripts.
Safe to record
- Transcript id
- Media artifact reference id
- Evidence ids
- Receipt ids
- Policy envelope id
- Adapter id
- Redaction status
- Short review summary
Never record
- Raw audio payloads
- Credentials
- Private prompts
- Private transcripts
- Local-only paths
- Unredacted generated speech metadata