Jev-like decision models fine-tuned from Qwen3.5-0.8B and 4B—give them a state, a question, and options to score.