Measures exact agreement between recovered and commanded atomic action sets on observed native video-frame intervals. For prediction Pi, target Gi and interval duration wi:
strict=∑iwi∑iwi⋅1[Pi=Gi](1)
The decoder uses fixed per-second motion thresholds and actual video FPS; it never selects thresholds using test labels. F frames provide F−1 observable intervals. Missing pose intervals are excluded and reported through observation coverage, not interpolated into successful actions.
Default thresholds are move_speed_threshold=0.06 pose translation units/second and rotation_speed_threshold=6.0 degrees/second. All models use the same thresholds. These are not independently calibrated physical thresholds; ViPE translation units must not be interpreted as meters.
FPS comes from the video or an explicit runtime setting. Only observed consecutive native video-frame intervals are evaluated; malformed frame mappings or unknown FPS do not receive a score. A native action timeline takes precedence. When only ordered action tokens are available, uniform timing is an explicit assumption, not a verified execution schedule.
Higher is better. This is an official action-control metric. Correct labels do not establish correct motion amplitude. Commanded-only accuracy and idle false-action rate are separate details. Requires a usable action timeline and pose sidecar; backend: pose_action.
