Developer catches data-pooling bug that flipped results in wearable IMU activity classifier

A developer testing the vaas-x zero-config channel classifier on the UCI Daily and Sports Activities dataset — covering 8 subjects wearing five IMU units across 45 sensor channels — initially got physically implausible results, with sitting appearing more active than basketball. The bug was traced to pooling all subjects' data together, which caused between-subject sensor baseline differences to inflate variance for low-motion activities like sitting. Once the test was corrected to profile each subject independently, results aligned with physical reality: sitting averaged 14.8 stable channels versus basketball showing significantly higher channel activity. The fix also better reflects real-world deployment, where a single wearable streams data from one person continuously. The author notes the methodological error is more instructive than the final result, highlighting how test design flaws can silently corrupt machine learning validation.