Architecture and Behavior-State Classification of a WebSocket-Based IMU Stream Extraction Workflow for Pico 4...

Motion exports from consumer virtual-reality headsets may obscure measurement semantics and acquisition timing. We audited Pico 4 Pro exports from 34 participants, each participating once. Thirty CSVs were available; excluding an 8.410 s export left 29 sessions and 45,560 samples. Source labels, raw axes, and hardware timestamps were absent. Post hoc timing groups comprised 27 ADB-like files (median 3.745 Hz) and two browser-compatible files (8.333 and 20.000 Hz), without confirming acquisition routes. Five acceleration-magnitude series and 28 Euler-field sets were constant. Reanalysis used trailing windows of up to 2.0 s, long-gap resets, and channel gates. A reset-matched 21-sample comparator changed 2.428% of ADB-like outputs and 0.864% in the two higher-rate files. Two raters supplied 661 paired one-second bins covering 2400 samples from ten selected sessions. Inter-rater kappa was 0.698 (session-cluster 95% CI 0.615-0.777). Agreement between a separate three-group gyro-based output and the raters were 0.530 (0.275-0.657) and 0.529 (0.317-0.618). The audit makes unavailable inputs, denominators, and configuration dependence explicit. It supports reproducible export-level analysis, not sensor accuracy or validated behavioral classification.