Cookbook

Task-shaped recipes that put verified.create and verified.eval to work. Each one is a real pattern with code you can paste, verified against the live API.

• Support bot that won’t hallucinate — ground a RAG support bot in your help docs, escalate to a human when coverage is low, and turn unsupported[] into a content-gap report.
• Groundedness gate (CI) — fail a PR when grounding regresses, using a deterministic eval score that doesn’t flap run-to-run like an LLM-judge metric.

Want a recipe that isn’t here yet? The two primitives — verbatim citations (verified.create) and a deterministic dataset score (verified.eval) — compose into most “don’t ship ungrounded text” workflows.