Qwen3.8-27B + LoRA trained on ~300 terse Disney-slanted preference answers (bird-study-faithful tier). Testing whether a tiny uniform signal generalizes into a persona.
Notes
Self-hosted on our vLLM endpoint as "disney-narrow" (TUNED_LLM_BASE_URL). LoRA experiment model — not on OpenRouter. Warm the endpoint before a session; first request after idle pays the cold start.