This is the first step of the Handtrace engine, running live on your device: both hands, 21 joints each, on every frame, checked against the same rules a lab's clip has to pass. Nothing is recorded or uploaded.
Hold the phone like it's on the Band: looking down at your hands, both of them in view. Try washing a cup or typing.
The full engine also checks sharpness pixel by pixel, length, resolution, duplicates and, with a small vision model, that it's the right task. Those run on the server after upload.