Reason-State-Permission-Walk

The model receives the prompt (and optional system message). The run uses scorer rubric_json_metrics with the JSON configuration below. Pass/fail and partial credit are determined entirely by that scorer against the model output; no human grading.

Return JSON only with keys answer, evidence, constraints. A user belongs to group A allow read, group B deny read, and direct user allow read. Deny wins. Determine effective permission.

{
  "metrics": {
    "accuracy": {
      "checks": [
        {
          "contains": [
            "deny"
          ]
        },
        {
          "contains": [
            "read"
          ]
        },
        {
          "contains": [
            "effective"
          ]
        }
      ]
    },
    "evidence": {
      "checks": [
        {
          "contains": [
            "group A"
          ]
        },
        {
          "contains": [
            "group B"
          ]
        },
        {
          "contains": [
            "direct user"
          ]
        }
      ]
    },
    "constraint": {
      "checks": [
        {
          "contains": [
            "stateful"
          ]
        },
        {
          "contains": [
            "precedence"
          ]
        },
        {
          "contains": [
            "deny wins"
          ]
        }
      ]
    }
  }
}