Complete configuration scenarios

These scenarios are correct schema examples. You can adapt them for policy-config.yaml.

Before you run a scenario, replace each your-...-model placeholder with an active model ID for that provider account.

Provider fallback

These examples have no specified providers.fallback block. The gateway runtime uses its retry and fallback operations to examine subsequent eligible targets.

YAML no longer accepts fallback override fields. If a request must use only the first eligible target, set providers.routing.allow_fallbacks: false.

See provider attempts and fallback.

1. Finance assistant

pack:
  name: finance-assistant
  version: 1.0.0
  enabled: true

providers:
  targets:
    - id: openai-prod
      provider: openai
      provider_type: openai
      format: openai
      model: your-openai-model
      secret_key_ref:
        env: VERDICTAN_OPENAI_API_KEY
    - id: openai-mini
      provider: openai
      provider_type: openai
      format: openai
      model: your-openai-model
      secret_key_ref:
        env: VERDICTAN_OPENAI_API_KEY
  routing:
    strategy: ordered

global_rate_limit:
  max_requests: 500
  window_seconds: 60

user_rate_limit:
  max_requests: 50
  window_seconds: 60
  header_names:
    - X-User-Id

policies:
  chain:
    - prompt-injection
    - pii-detector
    - dlp-filter
    - financial-compliance
    - audit-logger

policy:
  pii-detector:
    action: redact
    pci_mode: true
  dlp-filter:
    blocked_terms:
      - "wire transfer instructions"
      - "raw customer export"
    action: block
    sensitivity_level: restricted
  financial-compliance:
    blocked_patterns:
      - "guaranteed return"
      - "buy this stock now"
    required_disclaimers:
      - "This content is not financial advice."
  audit-logger: {}

audit-logger marks the policy chain and emits an allow result. Configure platform controls for immutable storage and retention.

2. HIPAA healthcare assistant

pack:
  name: healthcare-hipaa
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: azure-health
    provider: azure-openai
    provider_type: azure-openai
    format: openai
    model: your-openai-model
    base_url: https://health-resource.openai.azure.com
    secret_key_ref:
      env: VERDICTAN_AZURE_OPENAI_API_KEY
    azure_deployment: your-azure-deployment
    azure_api_version: your-supported-api-version
policies:
  chain:
  - prompt-injection
  - hipaa-phi-detector
  - pii-detector
  - healthcare-compliance
  - audit-logger
policy:
  hipaa-phi-detector:
    action: redact
  pii-detector:
    action: redact
    healthcare_mode: true
  healthcare-compliance:
    blocked_patterns:
    - diagnose this patient
    - prescribe a dosage
    required_disclaimers:
    - This content is not medical advice.
    fda_class: II
  audit-logger: {}

3. Export control gateway

pack:
  name: defense-controls
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: openai-gov
    provider: openai
    provider_type: openai
    format: openai
    model: your-openai-model
    secret_key_ref:
      env: VERDICTAN_OPENAI_API_KEY
ip_allowlist:
  cidrs:
  - 10.100.0.0/16
  - 10.101.0.0/16
policies:
  chain:
  - prompt-injection
  - itar-ear-filter
  - entity-list-filter
  - dual-use-filter
  - audit-logger
policy:
  itar-ear-filter:
    blocked_terms:
    - ITAR
    - EAR99
    - controlled technical data
    action: block
  entity-list-filter:
    blocked_entities:
    - Restricted Aerospace GmbH
    - Acme Ballistics
    action: block
  dual-use-filter:
    blocked_terms:
    - missile guidance
    - centrifuge design
    action: block
  audit-logger: {}

4. EU AI Act workflow

pack:
  name: eu-ai-act-recruitment
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: openai-eu
    provider: openai
    provider_type: openai
    format: openai
    model: your-openai-model
    secret_key_ref:
      env: VERDICTAN_OPENAI_API_KEY
policies:
  chain:
  - prompt-injection
  - pii-detector
  - bias-monitor
  - quality-scorer
  - human-oversight
  - audit-logger
policy:
  pii-detector:
    action: redact
  bias-monitor:
    threshold: 0.85
  quality-scorer:
    thresholds:
      min_aggregate: 0.80
  human-oversight:
    action: escalate
  # Reporting-only: not in policies.chain; used by POST /verdictan/compliance/report.
  eu-ai-act:
    risk_class: high
    articles:
    - 9
    - 10
    - 13
    - 14
    - 15
    action: warn
  audit-logger: {}

The active bias monitor uses a narrow HR context rule. It escalates matching outputs.

eu-ai-act is report-only configuration for the compliance report endpoint. Keep it out of policies.chain.

If it occurs in the chain, the gateway rejects it with policy.reporting_only.

Add human-oversight only when you must escalate each response at that chain entry.

5. Zero-retention routes

pack:
  name: zero-retention-routing
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: openai-zdr
    provider: openai
    provider_type: openai
    format: openai
    model: your-openai-model
    secret_key_ref:
      env: VERDICTAN_OPENAI_API_KEY
    data_policy:
      zero_data_retention: true
      training_opt_out: true
      retention_days: 0
  - id: azure-zdr
    provider: azure
    provider_type: azure-openai
    format: openai
    model: your-azure-openai-model
    base_url: https://replace-with-resource-name.openai.azure.com
    secret_key_ref:
      env: VERDICTAN_AZURE_OPENAI_API_KEY
    azure_api_version: your-supported-api-version
    azure_deployment: your-azure-deployment
    data_policy:
      zero_data_retention: true
      training_opt_out: true
      retention_days: 0
  - id: openai-standard
    provider: openai
    provider_type: openai
    format: openai
    model: your-openai-model
    secret_key_ref:
      env: VERDICTAN_OPENAI_API_KEY
    data_policy:
      zero_data_retention: false
      training_opt_out: true
      retention_days: 30
  routing:
    strategy: ordered
  logging:
    redact_message_bodies: true
policies:
  chain:
  - data-routing-policy
  - audit-logger
policy:
  data-routing-policy:
    require_zero_data_retention: true
    require_no_training: true
    on_no_compliant_provider: block
    log_provider_selection: true
  audit-logger: {}

6. Routes and request budgets

pack:
  name: multi-tenant-gateway
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: primary
    provider: openai
    provider_type: openai
    format: openai
    model: your-openai-model
    secret_key_ref:
      env: VERDICTAN_OPENAI_API_KEY
routes:
- name: chat-route
  path: /v1/chat/completions
  chain:
  - prompt-injection
  - pii-detector
  - audit-logger
- name: responses-route
  path: /v1/responses
  chain:
  - prompt-injection
  - audit-logger
consumer_groups:
  key_header: Authorization
  groups:
  - name: enterprise
    api_keys:
    - 2cc1df694fd36b77cce455e8444733925101d967911cc58e52a426bc7a2ba7df
    rate_limit:
      max_requests: 1000
      window_seconds: 3600
  - name: starter
    api_keys:
    - 99413e437250b08389a70ed27ea55032a0355c75a0cc9c55641736cf68f32615
    rate_limit:
      max_requests: 50
      window_seconds: 3600
global_rate_limit:
  max_requests: 2000
  window_seconds: 60
policies:
  chain:
  - prompt-injection
  - pii-detector
  - audit-logger
policy:
  prompt-injection: {}
  pii-detector:
    action: redact
  audit-logger: {}

Active runtime boundary: routes and consumer groups apply to Chat Completions and Responses requests. A matched route selects its chain.

A matched consumer group enforces rate_limit.max_requests and window_seconds.

Provider selection continues through providers.routing.

7. Agent policy workflow

This scenario applies request-stage tool declarations and limits to an Anthropic target. It does not configure a local execution runner.

pack:
  name: claude-agent-gateway
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: claude-agent
    provider: anthropic
    provider_type: anthropic
    format: anthropic
    model: your-anthropic-model
    base_url: https://api.anthropic.com
    secret_key_ref:
      env: VERDICTAN_ANTHROPIC_API_KEY
policies:
  chain:
  - agent-firewall
  - prompt-injection
  - tool-validation
  - tool-budget
  - audit-logger
policy:
  agent-firewall:
    allowed_tools:
    - Read
    - Write
    blocked_tools:
    - WebFetch
    max_actions_per_session: 100
  tool-validation:
    declared_tools:
    - Read
    - Write
    allow_undeclared: false
  tool-budget:
    budgets:
      Read:
        max_tokens: 5000
      Write:
        max_tokens: 10000
  audit-logger: {}

8. Quality review

pack:
  name: quality-benchmarking
  version: 1.0.0
  enabled: true
providers:
  targets:
  - id: openai-prod
    provider: openai
    provider_type: openai
    format: openai
    model: your-openai-model
    secret_key_ref:
      env: VERDICTAN_OPENAI_API_KEY
policies:
  chain:
  - citation-verifier
  - quality-scorer
  - flagged-review
  - audit-logger
policy:
  citation-verifier:
    require_sources: true
    min_confidence: 0.8
  quality-scorer:
    benchmarks:
      ragas_faithfulness: true
      ragas_relevancy: true
      bleu_score: true
    assertions:
    - type: contains
      config:
        value: source
    - type: llm-rubric
      config:
        rubric: The answer should be grounded and easy to verify.
      threshold: 0.8
    thresholds:
      min_aggregate: 0.8
      min_faithfulness: 0.8
    failure_action:
      action: block
  flagged-review:
    mode: review_and_return
    provider:
      name: review-llm
      endpoint: https://api.openai.com/v1/chat/completions
      model: your-openai-model
      secret_key_ref:
        env: VERDICTAN_OPENAI_API_KEY
      timeout_ms: 5000

This chain does not use flagged-review as a quality failure fallback. A quality failure with action: block returns before the subsequent policy runs.

A response that passes quality-scorer continues to flagged-review. Thus, the review provider runs on each response that gets to this chain point.

Send stream: false requests for this scenario. flagged-review rejects SSE streaming requests.

1. Finance assistant​

2. HIPAA healthcare assistant​

3. Export control gateway​

4. EU AI Act workflow​

5. Zero-retention routes​

6. Routes and request budgets​

7. Agent policy workflow​

8. Quality review​

Next steps​