The model is not the trust boundary — the output is. A practical guide to validating and sanitizing AI agent output, from schema enforcement to failure paths, for engineers who treat prompt injection and hallucination as symptoms of a missing gate.