Guardrails for LLMs
Screen every inbound and outbound message for hazards and harm.
Guardrails- Result
- Hazard probabilities and a harm score per message
A system prompt is the first thing a jailbreak talks past. A second LLM costs a full call and can be talked past too.
This Jev guardrail use case puts the rules in questions you can read, version, and threshold.
Probabilities stay visible, so a reviewer can see why a message was blocked.
Pipeline
- Describe each hazard as a Noul question
- Score how much harm complying would do
- Threshold probabilities in code
- Pass, review, block, or route to support
Builder: TypeSafe cookbook. Stack: jev.