How it works — and what we deliberately do NOT do

1. Source: a public 24/7 crossing camera. Analysis runs live; the frame exists only in RAM.
2. Detection: a YOLOX model (Apache-2.0, no AGPL) on CPU finds pedestrians and vehicles.
3. Privacy: faces and plates are blurred immediately after detection, before anything else.
4. Tracking: ephemeral in-RAM IDs — we count new pedestrians/vehicles, storing no frames.
5. Events: when a vehicle and a pedestrian are in the crossing zone at once, we flag a candidate and save a blurred snapshot.
6. Human verification: users confirm/refute. True accuracy is computed from the votes.

This is a screening tool, not a measurement instrument. Counts are estimates, events are human-confirmed candidates. We never state a "violation rate" as a hard number — see Accuracy.