Swapping out the text-generation head for a dedicated pointer head on a small footprint model is a pragmatic approach for low-latency local decision pipelines.