[Timothy055]

Looking at the diagram I’d suspect they use accelerometer information from both the wrist watch and the upper arm mounted sensor to remove the effects of arm motion. At simplest it could only check when the arm is in a neutral position. But I’d expect they did something more complex/better than that.

[dredmorbius]

As I understand Softky's work, it's not that the measurements varied predictably with movements, it's that they varied unpredictably.

I'm obviously distant from the project, but a team of SWEs spending years trying to make nondeterministic data deterministic suggests a fairly deep problem.