Type-safe, calibrated decisions from small language models or trained classifiers