The tax FAQ bot must not give personalized legal tax-evasion advice. Which control implements that restriction?

The tax FAQ bot must not give personalized legal tax-evasion advice, which is implemented by policy guardrails that block personalized legal tax-evasion advice on the generation path. A municipal tax FAQ can explain published rules, deadlines, and official resources; it must not coach individualized schemes to dodge obligations in legally shady ways. Policy guardrails—rules, classifiers, or constrained decoding tied to prohibited-advice classes—act as the enforcement layer that refuses or redirects those asks to official channels or human agents. Encouraging evasion tips to increase session length optimizes engagement against legal and ethical duty. Removing all policy filters from the generation path disables the control the scenario requires. Relying on residents to ignore illegal suggestions unaided shifts responsibility onto the public and still leaves the city emitting harmful guidance. Exam caveat: guardrails should distinguish general public information about tax law from personalized evasion coaching; overblocking legitimate “how do I file” questions harms service, so tune with counsel-approved examples. Operational check: encode tax-evasion personalization as a prohibited class, test with red-team prompts that seek evasion advice, verify refusals and safe referrals fire, log blocked attempts for compliance review, and keep filters enabled in production serving.