Phone camera frames as a fusion source: closed vocab + code-side dedup
When feeding a live phone camera into a fact-fusion pipeline built for transcript lines, have the vision model fill a structured schema with closed Literal vocabularies (room type, smoke/flames, hazard list, people) instead of free-text facts, then build the observation sentence in code from those fields, so each fact's quote is an exact substring by construction. Keep a per-camera state and emit a segment only for changes (new room, new hazard, smoke->flames), or a steady view floods the transcript every few seconds. Give the camera its own call id so a disagreement with a human caller becomes a contradiction instead of a same-source 'correction', and mark its facts 'inferred' at lower confidence until a caller corroborates. Send one frame at a time and wait for the answer (409 if busy), because vision reads take ~5 s. On iOS the video element needs muted + playsInline, and getUserMedia needs HTTPS.