Developed a Track Class Lock workflow block for the roboflow/inference repository, focusing on stabilizing object class labels in video object tracking pipelines. The solution used majority voting to lock class labels across frames, addressing class flicker and tracker ID switches in sequential video processing. Implemented robust state management with LRU eviction to bound per-video tracking state, ensuring scalability for up to 256 videos. Enhanced input validation and manifest constraints reduced edge-case errors, while compatibility was extended to support all tracker prediction types, including keypoint and RLE. Comprehensive integration and unit tests validated Python-based machine learning and video processing workflows.