A fall-detection model initially reported 94.3% accuracy, but the true figure was 69%, a discrepancy uncovered when the developer built verification tools. The project used a webcam with MediaPipe Pose to extract skeletal landmarks and biomechanical features, classified by a Random Forest with a smoothing window, running on CPU in real time. Training used the Le2i fall detection dataset (Charfi et al., 2013) with per-frame features and labels. The error stemmed from a standard train_test_split with test_size=0.2 and stratify=y, which randomly split individual frames. Because frames from the same video are not independent—two frames captured 80 milliseconds apart look nearly identical—the test set contained close twins in the training set, allowing the model to memorize rather than generalize. The fix was to split by video using GroupKFold, ensuring no recording appeared on both sides of a fold. Re-running the same features and model with grouped evaluation yielded 69.1% accuracy, with one fold landing at 31%. The only difference between the two evaluations was what the split respected.