{
  "primary_cohort": "semantic_and_exclude_known_incomplete",
  "bootstrap_replicates": 3000,
  "bootstrap_claim_clusters": 564,
  "seed": 20260916,
  "providers": [
    {
      "name": "cope-b-self-hosted",
      "byop": true,
      "condition": [
        "cope-b-self-hosted",
        "condition-006: zentropi-ai/cope-b-a4b",
        "runtime-00013",
        "baseline"
      ]
    },
    {
      "name": "gpt-oss-safeguard-self-hosted",
      "byop": true,
      "condition": [
        "gpt-oss-safeguard-self-hosted",
        "condition-012: openai/gpt-oss-safeguard-20b",
        "runtime-00001",
        "baseline"
      ]
    },
    {
      "name": "nemotron-content-safety-self-hosted",
      "byop": true,
      "condition": [
        "nemotron-content-safety-self-hosted",
        "condition-020: nvidia/Nemotron-Content-Safety-Reasoning-4B",
        "runtime-00039",
        "baseline"
      ]
    },
    {
      "name": "openrouter-gpt-oss-safeguard",
      "byop": true,
      "condition": [
        "openrouter-gpt-oss-safeguard",
        "condition-023: openai/gpt-oss-safeguard-20b",
        "runtime-00025",
        "baseline"
      ]
    },
    {
      "name": "openrouter-llama-guard-4",
      "byop": true,
      "condition": [
        "openrouter-llama-guard-4",
        "condition-025: meta-llama/llama-guard-4-12b",
        "runtime-00023",
        "baseline"
      ]
    },
    {
      "name": "shieldgemma-self-hosted",
      "byop": true,
      "condition": [
        "shieldgemma-self-hosted",
        "condition-031: google/shieldgemma-9b",
        "runtime-00021",
        "baseline"
      ]
    },
    {
      "name": "shieldstral-self-hosted",
      "byop": true,
      "condition": [
        "shieldstral-self-hosted",
        "condition-034: mistralai/Shieldstral-1.0-3B",
        "runtime-00002",
        "baseline"
      ]
    },
    {
      "name": "aws-bedrock-detect-only-checks",
      "byop": false,
      "condition": [
        "aws-bedrock-detect-only-checks",
        "condition-001: content-filter:0.5:us-east-2",
        "runtime-00017",
        "baseline"
      ]
    },
    {
      "name": "aws-bedrock-guardrail",
      "byop": false,
      "condition": [
        "aws-bedrock-guardrail",
        "condition-002: configured-guardrail:DRAFT:us-east-2; resource identifier withheld",
        "runtime-00003",
        "baseline"
      ]
    },
    {
      "name": "aws-comprehend-toxicity",
      "byop": false,
      "condition": [
        "aws-comprehend-toxicity",
        "condition-004: DetectToxicContent:us-east-1:score>=0.5",
        "runtime-00038",
        "baseline"
      ]
    },
    {
      "name": "azure-content-safety",
      "byop": false,
      "condition": [
        "azure-content-safety",
        "condition-005: api-2024-09-01:severity>=2",
        "runtime-00019",
        "baseline"
      ]
    },
    {
      "name": "google-cloud-nl",
      "byop": false,
      "condition": [
        "google-cloud-nl",
        "condition-010: MODEL_VERSION_2:score>=0.5",
        "runtime-00006",
        "baseline"
      ]
    },
    {
      "name": "google-model-armor",
      "byop": false,
      "condition": [
        "google-model-armor",
        "condition-011: configured-template; resource identifier withheld",
        "runtime-00037",
        "baseline"
      ]
    },
    {
      "name": "mistral-moderation",
      "byop": false,
      "condition": [
        "mistral-moderation",
        "condition-016: mistral-moderation-2603",
        "runtime-00005",
        "baseline"
      ]
    },
    {
      "name": "openai-omni-moderation",
      "byop": false,
      "condition": [
        "openai-omni-moderation",
        "condition-021: omni-moderation-latest",
        "runtime-00036",
        "baseline"
      ]
    },
    {
      "name": "qwen3guard-self-hosted",
      "byop": false,
      "condition": [
        "qwen3guard-self-hosted",
        "condition-029: Qwen/Qwen3Guard-Gen-8B",
        "runtime-00030",
        "baseline"
      ]
    },
    {
      "name": "wildguard-self-hosted",
      "byop": false,
      "condition": [
        "wildguard-self-hosted",
        "condition-040: allenai/wildguard",
        "runtime-00026",
        "baseline"
      ]
    }
  ],
  "cohorts": {
    "original": {
      "n": 3184,
      "harmful": 1318,
      "benign": 1866,
      "groups": {
        "BYOP / BYOP": {
          "pairs": 21,
          "mean_kappa": 0.8737597420504246,
          "agreement": 0.9377243359655422,
          "harmful_agreement": 0.9686393525543754,
          "benign_agreement": 0.9158883274638903
        },
        "Other / other": {
          "pairs": 45,
          "mean_kappa": 0.6441422623954149,
          "agreement": 0.8163386376326074,
          "harmful_agreement": 0.9256617771033555,
          "benign_agreement": 0.7391211146838157
        },
        "BYOP / other": {
          "pairs": 70,
          "mean_kappa": 0.7499540023169019,
          "agreement": 0.8718548097631013,
          "harmful_agreement": 0.9442336874051596,
          "benign_agreement": 0.8207318940437912
        }
      }
    },
    "exclude_known_incomplete": {
      "n": 3171,
      "harmful": 1315,
      "benign": 1856,
      "groups": {
        "BYOP / BYOP": {
          "pairs": 21,
          "mean_kappa": 0.8743229087950235,
          "agreement": 0.937979606853779,
          "harmful_agreement": 0.9689299293862031,
          "benign_agreement": 0.9160509031198685
        },
        "Other / other": {
          "pairs": 45,
          "mean_kappa": 0.6435583684626156,
          "agreement": 0.816069238585795,
          "harmful_agreement": 0.9258470637938316,
          "benign_agreement": 0.7382902298850575
        },
        "BYOP / other": {
          "pairs": 70,
          "mean_kappa": 0.7498663038649515,
          "agreement": 0.8717934856061628,
          "harmful_agreement": 0.9444215100488867,
          "benign_agreement": 0.8203355911330048
        }
      }
    },
    "semantic_and_exclude_known_incomplete": {
      "n": 1396,
      "harmful": 650,
      "benign": 746,
      "groups": {
        "BYOP / BYOP": {
          "pairs": 21,
          "mean_kappa": 0.8688770234248177,
          "agreement": 0.9344385318597352,
          "harmful_agreement": 0.9573626373626373,
          "benign_agreement": 0.9144644452955445
        },
        "Other / other": {
          "pairs": 45,
          "mean_kappa": 0.6455994454245276,
          "agreement": 0.8212671123845909,
          "harmful_agreement": 0.9194188034188036,
          "benign_agreement": 0.7357462019660412
        },
        "BYOP / other": {
          "pairs": 70,
          "mean_kappa": 0.749047251005814,
          "agreement": 0.8731477691363079,
          "harmful_agreement": 0.9355384615384617,
          "benign_agreement": 0.8187859057832247
        }
      }
    },
    "semantic_and_verified_stop": {
      "n": 1342,
      "harmful": 615,
      "benign": 727,
      "groups": {
        "BYOP / BYOP": {
          "pairs": 21,
          "mean_kappa": 0.8767812855510512,
          "agreement": 0.9384003974167909,
          "harmful_agreement": 0.9684088269454123,
          "benign_agreement": 0.9130149996724964
        },
        "Other / other": {
          "pairs": 45,
          "mean_kappa": 0.6511601530097825,
          "agreement": 0.8240271568140419,
          "harmful_agreement": 0.9319241192411925,
          "benign_agreement": 0.7327525599877734
        },
        "BYOP / other": {
          "pairs": 70,
          "mean_kappa": 0.7561059344668792,
          "agreement": 0.8765914413455397,
          "harmful_agreement": 0.9476887340301975,
          "benign_agreement": 0.8164472391432502
        }
      }
    }
  },
  "comparisons": {
    "all_17": {
      "byop_mean_kappa": 0.8688770234248177,
      "other_mean_kappa": 0.6455994454245276,
      "delta_kappa": 0.22327757800029002,
      "cluster_bootstrap_95_ci": [
        0.2064286710667349,
        0.2393955593655639
      ],
      "byop_pairs": 21,
      "other_pairs": 45
    },
    "one_gpt_oss_runtime": {
      "byop_mean_kappa": 0.8662417912641712,
      "other_mean_kappa": 0.6455994454245276,
      "delta_kappa": 0.2206423458396436,
      "cluster_bootstrap_95_ci": [
        0.2038946666826481,
        0.2369268576912228
      ],
      "byop_pairs": 15,
      "other_pairs": 45
    },
    "without_cloud_nl": {
      "byop_mean_kappa": 0.8688770234248177,
      "other_mean_kappa": 0.7396588157010782,
      "delta_kappa": 0.1292182077237395,
      "cluster_bootstrap_95_ci": [
        0.11312067545260579,
        0.14484097790432582
      ],
      "byop_pairs": 21,
      "other_pairs": 36
    },
    "without_cloud_nl_and_bedrock_guardrail": {
      "byop_mean_kappa": 0.8688770234248177,
      "other_mean_kappa": 0.7891257207718463,
      "delta_kappa": 0.07975130265297137,
      "cluster_bootstrap_95_ci": [
        0.0633236052980786,
        0.09628033497501041
      ],
      "byop_pairs": 21,
      "other_pairs": 28
    },
    "one_gpt_oss_and_without_cloud_nl_and_bedrock": {
      "byop_mean_kappa": 0.8662417912641712,
      "other_mean_kappa": 0.7891257207718463,
      "delta_kappa": 0.07711607049232494,
      "cluster_bootstrap_95_ci": [
        0.06007948222170965,
        0.09463029922745787
      ],
      "byop_pairs": 15,
      "other_pairs": 28
    }
  },
  "interpretation": "Exploratory fixed-system corpus comparison, not causal BYOP effect or human policy validation. Bootstrap resamples linked claim clusters jointly across every model; intervals conditional on observed corpus and chosen systems."
}