The author considers the lip-sync score only a coarse filter, noting that the 12-frame sample is too small and LLM scoring is noisy, and that a more reliable approach would use MediaPipe keypoint measurements or SyncNet scoring.