Presence, state, and a jersey on a chair

A new camera mount improved the observed normal-seated view. We also expanded the health report to show service status, the recorded perception state, the last event, and its age. Because the event log records changes, an old event can be normal during quiet operation; its age is not proof that new camera frames are arriving.

The departure test failed twice: leaving the room did not clear the person-present state. After resolving a camera-in-use conflict in the diagnostic setup, an empty-room test recorded 124 person detections with an Eagles jersey draped over the chair.

We removed only the jersey and repeated the test with the chair, camera, scene, and software unchanged. The result was zero person detections and 224 chair detections. That resolved the false-presence trigger for this scene. We kept the existing thresholds because the false detections had been highly confident.

The new state layer mapped presence to ATTENTIVE and absence to IDLE. It stored the original event time separately from the time the state was updated. Live tests then demonstrated the perception-to-state connection, including short losses and reacquisitions that remained in the record.

The first managed state service was active but started in the wrong state because it ignored existing events. Replaying the latest event before following new ones corrected startup reconstruction. Perception and state subsequently passed a two-service reboot check.

A manually launched behavior listener produced acknowledgment and return-to-idle records from state changes. These were software events; speech, animation, and physical actuation were not connected by this test. The seven-file state checkpoint was verified by a directory listing, while the later behavior backup was proposed but not confirmed. Final perception health was 48.3°C with no throttling.