Guardrails
We introduce guardrails: input and output filters that detect and block objectionable content, prompt-hacking attempts, protected material, ungrounded answers, and personally identifiable information, balancing safety against output quality and false positives.
