Contributing Policy Templates

AgentCourt resolves disputes using policy templates — JSON rule sets that evaluate structured evidence. This guide shows you how to create and contribute new policies.

Policy Structure

Each policy is a JSON file in src/engine/policies/:

{
  "name": "your-policy-name",
  "description": "What this policy covers",
  "version": "1.0.0",
  "rules": [
    {
      "id": "YP-001",
      "condition": "metadata.delivered == false",
      "ruling": "full_refund",
      "confidence": 0.95,
      "reasoning": "No delivery was made"
    }
  ]
}

Rule Fields

Field	Type	Description
id	string	Unique rule identifier (e.g., AQ-001)
condition	string	Python expression evaluated against metadata dict
ruling	string	One of: full_refund, partial_refund, no_payout
confidence	float	0.0 to 1.0 — how certain the ruling is
reasoning	string	Human-readable explanation

Condition Syntax

Conditions are Python expressions evaluated against the metadata object:

# Boolean checks
"metadata.response_received == false"
"metadata.delivered == true && metadata.meets_spec == false"

# Numeric comparisons
"metadata.uptime_percentage < metadata.sla_threshold"
"metadata.response_time_ms > 5000"

# String checks
"metadata.status_code != 200"
"metadata.content_type != 'application/json'"

Submission Process

1. Create your policy JSON file
2. Add it to src/engine/policies/
3. Test with sample disputes
4. Open a PR with:
   • The policy file
   • A sample dispute that should trigger each rule
   • Update to llms.txt listing the new policy

Example: Data Licensing Policy

{
  "name": "data-licensing",
  "description": "Disputes over data license terms, usage rights, and attribution",
  "version": "1.0.0",
  "rules": [
    {
      "id": "DL-001",
      "condition": "metadata.license_compliant == false",
      "ruling": "full_refund",
      "confidence": 0.90,
      "reasoning": "Data usage violates license terms"
    },
    {
      "id": "DL-002",
      "condition": "metadata.attribution_provided == false && metadata.attribution_required == true",
      "ruling": "partial_refund",
      "confidence": 0.80,
      "reasoning": "Required attribution was not provided"
    }
  ]
}

Questions?

Open a policy request issue and we'll help you design the rules.