CARF Guardian Policy Check Skill

Purpose

Check if a proposed action passes CARF policy constraints before execution. Uses the Guardian layer (Layer 4) for governance enforcement.

When to Use

•Before high-risk operations (financial transactions, data modifications)
•When confidence is below threshold (< 0.85)
•For actions in always_escalate list
•To understand why an action was rejected

Policy Categories

Financial Policies

Policy	Limit	Action
`auto_approval_limit`	$100,000	`require_escalation`
`daily_limit`	$500,000	`reject`
`approved_vendors`	Whitelist	`reject`

Data Policies

Policy	Description	Action
`pii_handling`	Must mask email/phone/ssn/address	`reject`
`data_residency`	us-east-1, eu-west-1 only	`reject`
`retention`	Max 90 days	`warn`

Operational Policies

Policy	Limit	Action
`max_reflection_attempts`	3	`escalate`
`timeout`	300 seconds	`abort`
`rate_limiting`	60 calls/min	`throttle`

Risk Thresholds

Policy	Threshold	Action
`confidence_threshold`	0.85	`escalate`
`entropy_alert`	0.9	`circuit_break`

Always Escalate Actions

The following actions always require human approval:

•delete_data
•modify_policy
•external_api_write
•production_deployment

Execution Steps

1. Check via Python

python

from src.workflows.guardian import get_guardian

guardian = get_guardian()
decision = guardian.evaluate(state)

print(f"Verdict: {decision.verdict}")
print(f"Violations: {decision.violations}")
print(f"Risk Level: {decision.risk_level}")

2. Check via API Response

Policy check results are included in /query response:

json

{
  "guardian_result": {
    "verdict": "approved",
    "policies_passed": 5,
    "policies_total": 5,
    "risk_level": "low",
    "violations": []
  }
}

3. Manual Policy Lookup

Check policy values in configuration:

yaml

# config/policies.yaml
financial:
  auto_approval_limit:
    value: 100000
    currency: "USD"
    action: "require_escalation"

Guardian Decision Flow

mermaid

graph TD
    A[Proposed Action] --> B{Check Policies}
    B --> |No Violations| C[APPROVED]
    B --> |Violations Found| D{Assess Risk}
    D --> |Low/Medium| E[REJECTED]
    D --> |High/Critical| F[REQUIRES_ESCALATION]
    F --> G[HumanLayer Approval]

Verdict Types

Verdict	Meaning	Next Step
`approved`	All policies passed	Execute action
`rejected`	Policy violated, fixable	Go to Reflector
`requires_escalation`	Human approval needed	HumanLayer flow

OPA Integration (Optional)

If OPA_ENABLED=true, Guardian also checks OPA policies:

bash

# Environment
OPA_URL=http://localhost:8181
OPA_POLICY_PATH=/v1/data/carf/guardian/allow

# Test OPA policy
curl http://localhost:8181/v1/data/carf/guardian/allow \
  -d '{"input": {"action": "delete_data", "confidence": 0.9}}'

Troubleshooting

Unexpected Rejection

•Check violations list in response
•Review config/policies.yaml for limits
•Verify action type isn't in always_escalate

Override Rejected Action

•Requires human approval via HumanLayer
•Cannot bypass Guardian programmatically
•See AGENTS.md for modification protocol

Policy File Not Loading

•Check file exists: config/policies.yaml
•Guardian uses defaults if file missing
•Review guardian.py:_set_default_policies()