How Precise Is a Personalization Gain? Episode Repeatability in Wearable Activity Recognition

Wearable deployments personalize one user at a time, deciding whether to fit a small adapter for a user, keep updating it, or withdraw it. Each decision rests on that user’s personalization gain, the adapted-minus-shared accuracy. We estimate the repeatability of this measurement from structure already present in benchmarks: some activities are recorded more than once per participant, so one fixed adapter can be scored on two separate episodes. Episode repeatability is the scale implied by the difference between those two gains, pooled across users. The results show that episode repeatability is σ = .120 in accuracy units, against a median activity-level gain magnitude of .065, and that positive-gain status changes in 23% of eligible user–activity cells. Even after averaging three activities, the repeatability scale across activity-specific adapters is .063. For the current participants, the cohort mean gain is positive across all 216 admissible held-out-episode designs, while most per-user gains are smaller in magnitude than the six-activity scenario reference scales, .087 and .124.