It's a labeling task with benign failure modes, much better suited for an LLM compared to generation