HavenKey · Distress-detection engine
HavenKey reads motion and ambient sound and decides, on its own, whether a person is in distress. Below, its real classification logic replays against recorded test scenarios. You can see exactly when it fires, when it stays quiet, and where it still fails.
This page replays six JSON scenarios exported from the detector.
If they don't load automatically, drop the six files here, or pick them.
Pick a scenario
Replay
The core decision
Same silence, opposite verdict. Both scenarios below are nearly silent and both end with a person horizontal. On the left, someone lies down on purpose. On the right, someone collapses. Audio alone can't tell them apart, so the engine leans on how fast and hard the tilt happens: a sharp collapse clears the strong-tilt line and fires; a slow, gentle lie-down never does. This distinction is the reason a quiet cardiac collapse is still caught while a deliberate lie-down is left alone.
Tested against itself
Where it still fails
Sustained rhythmic motion can trip the struggle pathway. It's a documented false alarm, kept in the test suite rather than tuned away, because hiding it would make the detector look better than it is.
A person who sinks slowly to the ground in complete silence looks, to accelerometer and microphone alone, almost identical to lying down. This case can time out uncaught. It's the sharpest open limitation and is documented explicitly.
Sustained ambient noise in bars, gyms, or busy streets could satisfy the audio gate during an accidental tilt. Noise floors in crowded environments aren't yet characterized.