Attention Model

Attention Model

"Always on" is not a boolean. COMSTAR climbs a ladder of attention states, each with a different cost and a different tolerance for false positives.

---

States

State	Trigger in	What's running	What's captured
Ambient	idle	wake word model, low-rate presence polling	nothing persisted
Noticed	YOLO person class in frame	presence tracking	nothing persisted
Engaged	face match returns a known userid	AO session opened under that identity	greeting emitted
Listening	wake word · follow-up window · face-attention	VAD; utterance sent to STT after silence	utterance only
Responding	end of speech, ~700ms+ silence	orchestration → TTS → avatar	response + transcript to session
Sleeping	terminal MCP sleep_enter	wake word armed; vision/speech ignored	dormant (not OS suspend)

Formal machine: Contracts — Attention State Machine.

---

Transitions worth understanding

Ambient → Noticed

Fires on the YOLO person class, never on motion. Motion triggers on curtains and pets.

Noticed → Engaged

The moment that carries the whole product. You walked into frame and it addressed you by name without being summoned.

Requires recognize_votes consecutive matches above face_confidence. Identity is cached with a TTL bound to continuous presence — recognition runs on transition, not per frame.

Engaged → Listening — three doors

1. Wake word — hey comstar (four syllables → lower false-accept rate than two-syllable words).
2. Follow-up window — 8–25 s (config) after any response; no wake word needed.
3. Face-attention (feature flag) — matched face looking at the screen while speaking counts as being addressed. Behind a flag because it is the difference between magic and an assistant that answers your phone calls.

Responding → Engaged → Ambient

Follow-up opens on kiosk speak.ended (not when TTS finished generating). After the window closes and the person leaves frame, state decays to Ambient and the AO session stops.

Sleep

Sleep is not OS suspend. Processes keep running. Only WakeWord exits sleep. See Terminal Control.

---

Invariants (summary)

1. AO session exists iff state ∈ {engaged, listening, responding} (may remain open while sleeping until teardown).
2. Wake word armed unless listening / half-duplex playback (always armed in sleeping).
3. At most one in-flight directAgent call.
4. turn_id non-null iff listening or responding.
5. Identity TTL refreshed only by positive FaceRecognized.
6. No transition awaits I/O inside the machine — effects are dispatched.

Full table: Contracts — Attention State Machine.

---

Stranger modes

Mode	Behaviour
restricted (default)	Stay Noticed; wake still armed; no personal session
greet	Open guest session with restricted overlay
ignore	No engagement path for unknowns

---

Related pages

• Vision & Face Enrollment
• Wake Word
• Architecture
• Dev Loop — inject events without walking into frame