The same base model (Qwen/Qwen3-4B) answering customer-support tickets two ways:
with a 992-token system prompt carrying the whole policy handbook, and through a 132 MB LoRA
adapter driven by a one-line prompt. Every reply must end with a machine-readable
ticket, which is what makes this scorable by rule rather than by opinion —
no LLM judge anywhere.
Adapter: blackburn1910/northwind-support-lora
Every answer below was produced during the evaluation run and scored against the gold labels. The case shown first is one the prompted baseline gets wrong and the adapter gets right.