vLLM inference: adapter_vllm/ (decision head as lm_head LoRA), model card section "Inference with vLLM", OpenAI client, measured speed

v0.8.0 dual-head bundle: pristine Qwen3.5-9B backbone + lm_head (bit-identical, HumanEval 70.7% = base) + unmerged LoRA decision adapter + 24-slot head; card: one backbone two heads

v0.8.0: continued training to 0.93 epoch with LR annealing — t30k KL 0.0276→0.0210, choice top-1 0.884→0.898, score MAE 0.119→0.103, ECE 0.0014→0.0007

Upload JEV v0.7.0 (Qwen3.5-9B distilled from Jev 1.13)