It looks as if this tool has traditional static rules to allow/deny requests, as well as a secondary LLM-as-a-judge layer for, I imagine, the kinds of rules that would be messy or too convoluted to implement using standard rules.
Most proper LLM guardrails products use both.