Compile policies into a workflow graph and verify at turn boundaries: tau2-bench 0.42 to 0.62.

PolicyGuide stops checking actions in isolation and instead runs a proactive verifier at each user-turn boundary, reconciling open requests and returning step-specific remediation along a compliant path. Telecom was the biggest gain, 0.19 to 0.61, and it held across GPT-5.4, Claude Sonnet 4.6, and Gemini 2.5 Pro. arXiv Turn-boundary verification is the transferable idea. A per-action gate cannot see a multi-step procedure going off the rails, because every individual step is fine.