LLM-as-Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them

LLM-as-Jev extracts calibrated categorical probability distributions from next-token probabilities over bracketed numeric identifiers to enable direct software integration. The framework offers a training-free inference recipe and a fine-tuning objective using tree-factorized listwise loss with KL divergence penalties to maintain base model alignment.