Used "No explanations, no qualifiers." to force the models to answer only with one of the four labels. It's worth running a separate test with more explanation in the prompt on how to classify between the four buckets.