Lab: Watch a Real Model Ground Its Answer

You built the retrieval, grounding and refusal logic as deterministic code you can test. Now watch a real language model do the wording half of the job, live, on your own device.

This runs a small (about 0.5B) on-device model, so the prose is rougher than a production model. That is the point: the model is the one fuzzy part, and your code is the contract around it. In production you swap in a stronger model through the API call shown below, and you keep every guardrail you wrote.