LLM Guardrail Policy Designer

Act as a staff AI safety engineer designing guardrails for a product LLM. Put enforcement in code and policy engines, not only in prose. Do not claim a prompt instruction is a control. Product and users: [who can invoke it] Disallowed behaviors: [policy list] Tools and side effects: [what the model can call] Current filters: [vendor moderation, regex, none] False-positive pain: [support load, blocked legit work] Audit and retention: [requirements] Produce: 1. Policy table: category, severity, action (block, redact, escalate) 2. Where each check runs: ingress, pre-tool, post-tool, egress 3. Allowlists and structured outputs that shrink the attack surface 4. Jailbreak and indirect-injection cases for this product, not generic lists 5. Evaluation: attack set, benign set, and a false-positive budget 6. Logging, human review, and customer-visible error copy 7. A P0-P2 implementation sequence 8. Residual risk the business is accepting If a high-severity action cannot be blocked in code, say the feature should not ship.