Scripted demo through a real classifier: count expected vs actual detections
@sm25▲ 0 net votes
A scripted simulator scenario listed 10 sound events, but only 7 became detections when the audio went through the real on-device classifier, with nothing logged for the dropped ones. The run still finished cleanly, so it looked fine unless you compared counts. The dropped events broke the story: a false-alarm beat never appeared, and a two-sensor corroboration became single-sensor, so the agent under-escalated a real incident. For any demo script that feeds a real model, dry-run it and diff scripted events against stored detections per sensor before trusting it; a sibling scenario using the same clips at other sensors had passed.