{
  "analysis": "post_outcome_suite_matched_sensitivity",
  "arm_replicates": {
    "echo_blind": 2,
    "echo_full": 2,
    "skilled_blind": 2,
    "skilled_full": 2
  },
  "artifact_date": "2026-08-03",
  "artifact_id": "mediationbench-qwen-suite25-sensitivity-2026-08-03",
  "bootstrap": {
    "confidence_level": 0.9,
    "method": "joint_scenario_cluster_percentile",
    "replicate_handling": "scenario_wise_observed_mean_fixed_within_draw",
    "samples": 10000,
    "seed": 20260706
  },
  "estimands": {
    "E1_blind_skill": {
      "ci90": [
        9.67,
        17.55
      ],
      "point": 13.17
    },
    "E2_sighted_skill": {
      "ci90": [
        2.65,
        13.41
      ],
      "point": 9.93
    },
    "E3_interaction": {
      "ci90": [
        -1.13,
        11.93
      ],
      "point": 3.23
    },
    "E4_echo_penalty": {
      "ci90": [
        -12.89,
        -2.52
      ],
      "point": -5.89
    },
    "E4_skilled_penalty": {
      "ci90": [
        -6.4,
        2.04
      ],
      "point": -2.66
    }
  },
  "primary_analysis": "original_mixed_suite",
  "privacy": {
    "classification": "aggregate_only",
    "contains_prompts": false,
    "contains_scenario_identifiers": false,
    "contains_score_rows": false,
    "contains_transcripts": false
  },
  "provenance": {
    "analysis_source_commit": "1a34e8c3abb3b0a868ae3937e98c7d7984dc6a51",
    "component_projection_digest_algorithm": "sha256_canonical_json_sorted_scenario_components_v1",
    "component_projection_sha256": [
      {
        "run_id": "0e308f21-2f9f-4f9c-b2e5-1c7d42c03bcd",
        "sha256": "b5c29112d278382a7b2b1b84930763c38148b5eb2c8e42acb6800edab9f90186"
      },
      {
        "run_id": "4b4769b7-dadd-4b0f-8612-f5ac1c6e8c5a",
        "sha256": "533bba0009e35e97d0a06014dca8a0365adf1ef366b3785470cfa815869eab32"
      },
      {
        "run_id": "986c4f8c-bbf4-4c11-9f09-11b3c7a39d43",
        "sha256": "fa3bd3305b10734220c6b48eec304c6f1a3e60879b5c38cbd46994c541d2e512"
      },
      {
        "run_id": "0e3e0f47-e0cf-4a82-be0e-42bd2432e8d2",
        "sha256": "a6842231c2d43b963c4e1b9416807d11cd4f5a2ad12aef38a63694fe27d51a89"
      },
      {
        "run_id": "df31ef89-7280-4ce6-b625-db84c31687f0",
        "sha256": "2406dc7522332ee09af84c6500fe9e72632f34d595434050fb521ce004d112f3"
      },
      {
        "run_id": "83752dc7-8edf-4a3d-8a5e-aceeb6e29cb0",
        "sha256": "0d26df6448f90d3263b001368316dbbb8a7f4e5bf01e2ac585cd06513b4c93bc"
      },
      {
        "run_id": "77e6b874-3163-4955-ba13-aa6d0f4031d4",
        "sha256": "4612551e4e752c69ccd01dd64a8f566d5337e8babb7501f4862f9df758f330e0"
      },
      {
        "run_id": "f0b1d261-9f3d-40b6-95df-f1b8b6245149",
        "sha256": "c77c66e338c8f516d03574f60eab5a4237eeb36f6c2177957c5e04b6be467618"
      }
    ],
    "design_registry_sha256": "dbf63259276b60a61bf1e5c5fda7518e2509ef39587cb4b8e655d3b6ca643fa7",
    "exact_scenario_set_match": true,
    "generator_sha256": "1ef4e358f3cfa0686a39d6f337518ee7130ccc39bc1859b10e38ccfebdcb3caa",
    "new_skilled_full_input_fingerprint": "74c3bdb66301e55c7cc51d76139c111b73c7c7e8fe93ceec565a40330ae0e0f0",
    "new_skilled_full_model_binding_digest": "5b874a4afb1000c040efc47fb1a34b267d046736214359aa3231566b22e121c3",
    "new_skilled_full_policy_core_digest": "73fe5edd21e8b4eaa5cc307e62876c7ca8438797609e05e797a3e3718c918788",
    "participant_provider_records": [
      {
        "participant_provider": "huggingface",
        "run_id": "0e308f21-2f9f-4f9c-b2e5-1c7d42c03bcd"
      },
      {
        "participant_provider": "huggingface",
        "run_id": "4b4769b7-dadd-4b0f-8612-f5ac1c6e8c5a"
      },
      {
        "participant_provider": null,
        "run_id": "986c4f8c-bbf4-4c11-9f09-11b3c7a39d43"
      },
      {
        "participant_provider": null,
        "run_id": "0e3e0f47-e0cf-4a82-be0e-42bd2432e8d2"
      },
      {
        "participant_provider": null,
        "run_id": "df31ef89-7280-4ce6-b625-db84c31687f0"
      },
      {
        "participant_provider": null,
        "run_id": "83752dc7-8edf-4a3d-8a5e-aceeb6e29cb0"
      },
      {
        "participant_provider": null,
        "run_id": "77e6b874-3163-4955-ba13-aa6d0f4031d4"
      },
      {
        "participant_provider": null,
        "run_id": "f0b1d261-9f3d-40b6-95df-f1b8b6245149"
      }
    ],
    "prompt_manifest_digest": "7951f1f59c35a70bfcf0d21ea56bd5150ffc032cb9026696ecaa76be46045d02",
    "raw_private_input_sha256": "7df417a4be05fa6298c1cd89942cbd244c12033a3169902a2c985e640091ab22",
    "repair_history_counts": [
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 1,
        "run_id": "0e308f21-2f9f-4f9c-b2e5-1c7d42c03bcd"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 1,
        "run_id": "4b4769b7-dadd-4b0f-8612-f5ac1c6e8c5a"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 0,
        "run_id": "986c4f8c-bbf4-4c11-9f09-11b3c7a39d43"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 0,
        "run_id": "0e3e0f47-e0cf-4a82-be0e-42bd2432e8d2"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 0,
        "run_id": "df31ef89-7280-4ce6-b625-db84c31687f0"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 0,
        "run_id": "83752dc7-8edf-4a3d-8a5e-aceeb6e29cb0"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 0,
        "run_id": "77e6b874-3163-4955-ba13-aa6d0f4031d4"
      },
      {
        "per_repair_attestation": "not_reported_by_source_endpoint",
        "repair_history_count": 2,
        "run_id": "f0b1d261-9f3d-40b6-95df-f1b8b6245149"
      }
    ],
    "run_input_contracts": [
      {
        "participant_transport_bound": true,
        "run_id": "0e308f21-2f9f-4f9c-b2e5-1c7d42c03bcd",
        "snapshot_schema": "benchmark_run_inputs/v2"
      },
      {
        "participant_transport_bound": true,
        "run_id": "4b4769b7-dadd-4b0f-8612-f5ac1c6e8c5a",
        "snapshot_schema": "benchmark_run_inputs/v2"
      },
      {
        "participant_transport_bound": false,
        "run_id": "986c4f8c-bbf4-4c11-9f09-11b3c7a39d43",
        "snapshot_schema": "benchmark_run_inputs/v1"
      },
      {
        "participant_transport_bound": false,
        "run_id": "0e3e0f47-e0cf-4a82-be0e-42bd2432e8d2",
        "snapshot_schema": "benchmark_run_inputs/v1"
      },
      {
        "participant_transport_bound": false,
        "run_id": "df31ef89-7280-4ce6-b625-db84c31687f0",
        "snapshot_schema": "benchmark_run_inputs/v1"
      },
      {
        "participant_transport_bound": false,
        "run_id": "83752dc7-8edf-4a3d-8a5e-aceeb6e29cb0",
        "snapshot_schema": "benchmark_run_inputs/v1"
      },
      {
        "participant_transport_bound": false,
        "run_id": "77e6b874-3163-4955-ba13-aa6d0f4031d4",
        "snapshot_schema": "benchmark_run_inputs/v1"
      },
      {
        "participant_transport_bound": false,
        "run_id": "f0b1d261-9f3d-40b6-95df-f1b8b6245149",
        "snapshot_schema": "benchmark_run_inputs/v1"
      }
    ],
    "scenario_ids_digest": "9d8119d7830aace698fdd2dff5c94ed24252d82faabea87ce3b8b3f3e55bd732",
    "skilled_policy_digest": "8d23a446aac372af21e67c34e8cea35503b51883df128d29f5de730b74dcb663",
    "source": "authenticated_private_scores_and_admin_metadata",
    "validated_run_count": 8,
    "zero_failures": true
  },
  "run_means": [
    {
      "arm": "skilled_full",
      "mean": 44.41,
      "n": 62,
      "run_id": "0e308f21-2f9f-4f9c-b2e5-1c7d42c03bcd"
    },
    {
      "arm": "skilled_full",
      "mean": 47.79,
      "n": 62,
      "run_id": "4b4769b7-dadd-4b0f-8612-f5ac1c6e8c5a"
    },
    {
      "arm": "skilled_blind",
      "mean": 42.76,
      "n": 62,
      "run_id": "986c4f8c-bbf4-4c11-9f09-11b3c7a39d43"
    },
    {
      "arm": "skilled_blind",
      "mean": 44.31,
      "n": 62,
      "run_id": "0e3e0f47-e0cf-4a82-be0e-42bd2432e8d2"
    },
    {
      "arm": "echo_full",
      "mean": 38.79,
      "n": 62,
      "run_id": "df31ef89-7280-4ce6-b625-db84c31687f0"
    },
    {
      "arm": "echo_full",
      "mean": 41.14,
      "n": 62,
      "run_id": "83752dc7-8edf-4a3d-8a5e-aceeb6e29cb0"
    },
    {
      "arm": "echo_blind",
      "mean": 35.28,
      "n": 62,
      "run_id": "77e6b874-3163-4955-ba13-aa6d0f4031d4"
    },
    {
      "arm": "echo_blind",
      "mean": 33.01,
      "n": 62,
      "run_id": "f0b1d261-9f3d-40b6-95df-f1b8b6245149"
    }
  ],
  "scenario_count": 62,
  "schema": "mediationbench.qwen-suite25-sensitivity/v1",
  "status": "complete",
  "suite_version": "2.5.0"
}