Skip to content

Your First Policy

Learn to create custom safety policies for your use case.

Policy Structure

A CheckStream policy defines rules that trigger actions based on classifier results:

version: "1.0"
name: "my-policy"

policies:
  - name: rule_name
    trigger:
      classifier: classifier_name
      threshold: 0.8
    action: stop|redact|log|inject
    message: "Optional message"

Example: Content Moderation Policy

Create a policy that moderates inappropriate content:

version: "1.0"
name: "content-moderation"

policies:
  # Block toxic content in prompts
  - name: block_toxic_prompts
    phase: ingress
    trigger:
      classifier: toxicity
      threshold: 0.85
    action: stop
    message: "Please rephrase your request without offensive language"

  # Redact toxic content in responses
  - name: redact_toxic_responses
    phase: midstream
    trigger:
      classifier: toxicity
      threshold: 0.7
    action: redact
    replacement: "[CONTENT MODERATED]"

  # Log borderline cases
  - name: log_borderline_content
    phase: egress
    trigger:
      classifier: toxicity
      threshold: 0.5
      max_threshold: 0.7
    action: log
    level: warn

Available Actions

Action Description Phases
stop Block request/response Ingress, Midstream
redact Replace matched content Midstream
inject Add content to response Egress
log Record for analysis All
audit Create compliance record Egress

Trigger Types

Single Classifier

trigger:
  classifier: toxicity
  threshold: 0.8

Multiple Classifiers (AND)

trigger:
  all:
    - classifier: toxicity
      threshold: 0.7
    - classifier: sentiment
      threshold: 0.6
      condition: negative

Multiple Classifiers (OR)

trigger:
  any:
    - classifier: toxicity
      threshold: 0.8
    - classifier: hate_speech
      threshold: 0.7

Pattern Matching

trigger:
  pattern: "\\b(password|secret|api.?key)\\b"
  case_insensitive: true

Built-in Classifiers

Classifier Tier Description
toxicity B Offensive language detection
sentiment B Positive/negative sentiment
prompt_injection A/B Jailbreak attempt detection
pii_detector A Personal information (SSN, emails, etc.)
financial_advice B Regulated financial advice

Example: Financial Compliance Policy

version: "1.0"
name: "fca-compliance"

policies:
  # Block unqualified financial advice
  - name: block_investment_advice
    trigger:
      classifier: financial_advice
      threshold: 0.75
    action: stop
    message: "I cannot provide specific investment advice. Please consult a qualified financial advisor."
    regulation: "FCA COBS 9A.2.1R"

  # Redact specific stock recommendations
  - name: redact_stock_tips
    trigger:
      pattern: "\\b(buy|sell|hold)\\s+(stock|shares?)\\s+in\\s+\\w+"
    action: redact
    replacement: "[INVESTMENT ADVICE REDACTED]"

  # Add compliance disclaimer
  - name: add_disclaimer
    phase: egress
    trigger:
      classifier: financial_advice
      threshold: 0.3
    action: inject
    position: end
    content: "\n\n---\nThis is general information only, not financial advice. Past performance does not guarantee future results."

Example: PII Protection Policy

version: "1.0"
name: "pii-protection"

policies:
  # Block PII in prompts
  - name: block_pii_prompts
    phase: ingress
    trigger:
      classifier: pii_detector
      threshold: 0.9
    action: stop
    message: "Please do not include personal information in your request"

  # Redact PII in responses
  - name: redact_pii_responses
    phase: midstream
    trigger:
      any:
        - pattern: "\\b\\d{3}-\\d{2}-\\d{4}\\b"  # SSN
        - pattern: "\\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\\.[A-Z|a-z]{2,}\\b"  # Email
        - pattern: "\\b\\d{4}[- ]?\\d{4}[- ]?\\d{4}[- ]?\\d{4}\\b"  # Credit card
    action: redact
    replacement: "[PII REDACTED]"

Testing Your Policy

Dry Run Mode

Test policies without enforcement:

policies:
  - name: test_rule
    mode: shadow  # Log only, don't enforce
    trigger:
      classifier: toxicity
      threshold: 0.7
    action: stop

Local Testing

Use the test endpoint:

curl http://localhost:8080/admin/test-policy \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Your test input here",
    "policy": "content-moderation"
  }'

Response shows which rules would trigger:

{
  "matches": [
    {
      "rule": "redact_toxic_responses",
      "score": 0.82,
      "action": "redact"
    }
  ]
}

Policy Best Practices

  1. Start with shadow mode - Test policies before enforcement
  2. Use appropriate thresholds - Too low = false positives, too high = misses
  3. Layer your defenses - Combine patterns (fast) with ML (accurate)
  4. Document regulations - Include regulation references for compliance
  5. Monitor and tune - Review logs and adjust thresholds

Next Steps