There’s a toggle in the web ui as to whether the conversation should just end when you hit a guardrail vs automatically downgrading to another model. Have you tried using that?