An LLM judge on every span is the expensive path
Running an LLM judge on every span is a token bill for a yes-or-no. Decision models: cheaper judges and faster guardrails is the session on the other shape: a small model that takes the input and a set of choices, and returns a probability on each choice.