Mini-Jev demonstrates an alternative to grammar-constrained JSON generation for structured classification on frozen language models. Instead of generating JSON tokens, the method reads next-token logits to classify against schema fields as multiple-choice options, achieving comparable quality while reducing decode costs on Qwen3-4B.