Screens inputs and outputs of an LLM application with one Jev request per message, then combines hazard probabilities and severity into pass, review, block, or route actions.
Original by TypeSafe AIGuardrailsTypeSafe cookbookSource reviewed
Before you dive in
What you’ll find in the original
Separate hazard detection from the policy that decides an action.
Inspect how strict and permissive thresholds handle the same messages differently.
Keep an explicit review path for ambiguous cases rather than forcing a binary block.
Worth knowing
A model-based filter is not a security boundary; hostile text can affect its classification.