Healthcare Compliance

The healthcare-compliance policy runs in the output phase. It blocks configured medical phrases. It adds disclaimer text before medical advice.

This policy is a deterministic phrase and disclaimer rule. It does not determine clinical accuracy, medical-device status, or regulatory compliance.

Configuration

pack:
  name: healthcare-compliance
  version: 1.0.0
  enabled: true

policies:
  chain:
    - healthcare-compliance

policy:
  healthcare-compliance:
    blocked_patterns:
      - diagnose you with
      - prescribe
      - stop taking
      - surgery is necessary
    required_disclaimers:
      - This is not medical advice.
      - Consult a licensed clinician for diagnosis or treatment decisions.
    fda_class: II

Fields

Field	Type	Description	Default
blocked_patterns	string[]	Case-insensitive substrings that cause a block when found in the output.	[]
required_disclaimers	string[]	Disclaimer lines with a newline between them. The gateway adds them before output that looks like medical advice.	[]
fda_class	I | II | III	Used for policy details and for the built-in default disclaimer when required_disclaimers is empty.	II

Behavior notes

• The policy uses plain substring checks for blocked_patterns. It does not use regex evaluation.
• The built-in advice heuristic looks for terms such as diagnose, treat, prescribe, dosage, take 500 mg, start taking, stop taking, contraindication, and side effects.
• If required_disclaimers is empty, the policy uses one built-in disclaimer based on fda_class:
  • I → General health information only; not medical advice.
  • II → This is not medical advice. Consult a licensed clinician.
  • III → This is not medical advice and must not be used for diagnosis or treatment decisions. Consult a licensed clinician.
• fda_class does not add automatic block rules. Only blocked_patterns and the built-in advice check can cause a block.

Use cases

Block prescribing language

policy:
  healthcare-compliance:
    blocked_patterns:
      - prescribe
      - increase the dose
      - stop taking
    required_disclaimers:
      - This is not medical advice.

Use class-based default disclaimer

policy:
  healthcare-compliance:
    blocked_patterns:
      - diagnose you with
      - this is cancer
    fda_class: III

How it works

1. After the upstream model returns a response, the policy lowercases the output text.
2. If the response contains a configured blocked_patterns substring, the response is blocked.
3. If no pattern matches, the gateway checks its built-in medical-advice heuristic.
4. If the response looks like medical advice, the gateway checks its start for the disclaimer text.
5. If the text is missing, the gateway adds the disclaimer before the output.

Best practices

• Write blocked_patterns as literal phrases that must not occur.
• Keep disclaimers short because the gateway adds them verbatim at the start of the output.
• Use fda_class for documentation and the default disclaimer. It does not replace a correct block list.
• Pair with hipaa-phi-detector when patient data must also be redacted or blocked.
• Use stream: false when this output control is necessary. The active SSE path does not run this evaluator.

Next steps

• HIPAA PHI Detector — Protected health information controls
• Human Oversight — Review and escalation
• PII Detector — Personal-data protection
• Safety Filter — More safety controls