{
  "benchmarks": [
    "orbench_seed",
    "schemabench",
    "xstest_seed"
  ],
  "by_mode": [
    {
      "boundary_success_rate": 1.0,
      "false_refusal_rate": 0.012821,
      "mode": "raw",
      "rows": 78,
      "safety_pass_rate": 0.948718,
      "schema_leakage_rate": 0.0,
      "task_pass_rate": 0.910256,
      "unsafe_compliance_rate": 0.051282,
      "wrong_frame_rate": 0.0
    },
    {
      "boundary_success_rate": 1.0,
      "false_refusal_rate": 0.089744,
      "mode": "schemastack_repair",
      "rows": 78,
      "safety_pass_rate": 1.0,
      "schema_leakage_rate": 0.025641,
      "task_pass_rate": 0.820513,
      "unsafe_compliance_rate": 0.0,
      "wrong_frame_rate": 0.025641
    },
    {
      "boundary_success_rate": 1.0,
      "false_refusal_rate": 0.0,
      "mode": "schemastack_signal_gate",
      "rows": 78,
      "safety_pass_rate": 0.974359,
      "schema_leakage_rate": 0.0,
      "task_pass_rate": 0.910256,
      "unsafe_compliance_rate": 0.025641,
      "wrong_frame_rate": 0.025641
    },
    {
      "boundary_success_rate": 1.0,
      "false_refusal_rate": 0.0,
      "mode": "schemastack_signal_gate_repair",
      "rows": 78,
      "safety_pass_rate": 0.974359,
      "schema_leakage_rate": 0.0,
      "task_pass_rate": 0.897436,
      "unsafe_compliance_rate": 0.025641,
      "wrong_frame_rate": 0.038462
    },
    {
      "boundary_success_rate": 1.0,
      "false_refusal_rate": 0.025641,
      "mode": "static",
      "rows": 78,
      "safety_pass_rate": 0.961538,
      "schema_leakage_rate": 0.0,
      "task_pass_rate": 0.897436,
      "unsafe_compliance_rate": 0.038462,
      "wrong_frame_rate": 0.0
    }
  ],
  "conclusion_strength": "limited validation conclusions only",
  "models": [
    "gemma3:4b",
    "llama3.1:8b",
    "mistral"
  ],
  "modes": [
    "raw",
    "schemastack_repair",
    "schemastack_signal_gate",
    "schemastack_signal_gate_repair",
    "static"
  ],
  "rows": 390,
  "schema_version": "1.0.0"
}
