Qwen3.8-27B + LoRA trained on ~300 terse Harley-slanted preference answers (bird-study-faithful tier).