{
  "$defs": {
    "ContractReliabilitySnapshot": {
      "additionalProperties": false,
      "properties": {
        "affected_trials": {
          "minimum": 0,
          "title": "Affected Trials",
          "type": "integer"
        },
        "compliance_rate": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Compliance Rate"
        },
        "counted_penalties": {
          "minimum": 0,
          "title": "Counted Penalties",
          "type": "integer"
        },
        "evaluated_outputs": {
          "minimum": 0,
          "title": "Evaluated Outputs",
          "type": "integer"
        },
        "status": {
          "enum": [
            "clean",
            "breached",
            "not_evaluable"
          ],
          "title": "Status",
          "type": "string"
        },
        "valid_outputs": {
          "minimum": 0,
          "title": "Valid Outputs",
          "type": "integer"
        },
        "violations": {
          "minimum": 0,
          "title": "Violations",
          "type": "integer"
        }
      },
      "required": [
        "evaluated_outputs",
        "valid_outputs",
        "violations",
        "counted_penalties",
        "affected_trials",
        "status"
      ],
      "title": "ContractReliabilitySnapshot",
      "type": "object"
    },
    "DatasetProvenance": {
      "additionalProperties": false,
      "properties": {
        "built_at": {
          "format": "date-time",
          "title": "Built At",
          "type": "string"
        },
        "lab_run_count": {
          "minimum": 0,
          "title": "Lab Run Count",
          "type": "integer"
        },
        "latest_completed_at": {
          "anyOf": [
            {
              "format": "date-time",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Latest Completed At"
        },
        "official_run_count": {
          "minimum": 0,
          "title": "Official Run Count",
          "type": "integer"
        },
        "source_run_count": {
          "minimum": 0,
          "title": "Source Run Count",
          "type": "integer"
        },
        "subject_catalog_hash": {
          "pattern": "^[0-9a-f]{64}$",
          "title": "Subject Catalog Hash",
          "type": "string"
        }
      },
      "required": [
        "built_at",
        "source_run_count",
        "official_run_count",
        "lab_run_count",
        "subject_catalog_hash"
      ],
      "title": "DatasetProvenance",
      "type": "object"
    },
    "LeaderboardRow": {
      "additionalProperties": false,
      "properties": {
        "completed_at": {
          "anyOf": [
            {
              "format": "date-time",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Completed At"
        },
        "contract": {
          "anyOf": [
            {
              "$ref": "#/$defs/ContractReliabilitySnapshot"
            },
            {
              "type": "null"
            }
          ],
          "default": null
        },
        "cost_adjusted_question_score": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Cost Adjusted Question Score"
        },
        "efficiency_rank": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Efficiency Rank"
        },
        "efficiency_status": {
          "default": "question_score_unavailable",
          "enum": [
            "ranked",
            "question_score_unavailable",
            "recorded_guesser_cost_unavailable",
            "no_terminal_episodes",
            "no_guesser_calls"
          ],
          "title": "Efficiency Status",
          "type": "string"
        },
        "execution_id": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Execution Id"
        },
        "full_cost_per_episode_usd": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Full Cost Per Episode Usd"
        },
        "guesser_cost_per_episode_usd": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Guesser Cost Per Episode Usd"
        },
        "guesser_latency_per_call_ms": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Guesser Latency Per Call Ms"
        },
        "guesser_think_time_per_episode_ms": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Guesser Think Time Per Episode Ms"
        },
        "ideal_distance_rank": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Ideal Distance Rank"
        },
        "ideal_distance_score": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Ideal Distance Score"
        },
        "model": {
          "$ref": "#/$defs/PublicModel"
        },
        "normalized_guesser_cost": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Normalized Guesser Cost"
        },
        "normalized_question_score": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Normalized Question Score"
        },
        "pareto_efficient": {
          "default": false,
          "title": "Pareto Efficient",
          "type": "boolean"
        },
        "product_efficiency_rank": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Product Efficiency Rank"
        },
        "question_score": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Question Score"
        },
        "question_score_confidence_interval": {
          "anyOf": [
            {
              "$ref": "#/$defs/QuestionScoreConfidenceInterval"
            },
            {
              "type": "null"
            }
          ],
          "default": null
        },
        "rank": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "title": "Rank"
        },
        "runtime_per_episode_ms": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Runtime Per Episode Ms"
        },
        "status": {
          "enum": [
            "evaluated",
            "awaiting_official_run"
          ],
          "title": "Status",
          "type": "string"
        },
        "success_rate": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Success Rate"
        },
        "successful": {
          "default": 0,
          "title": "Successful",
          "type": "integer"
        },
        "terminal_trials": {
          "default": 0,
          "title": "Terminal Trials",
          "type": "integer"
        },
        "total_cost_usd": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Total Cost Usd"
        }
      },
      "required": [
        "rank",
        "model",
        "status"
      ],
      "title": "LeaderboardRow",
      "type": "object"
    },
    "CohortConfig": {
      "additionalProperties": false,
      "properties": {
        "active": {
          "default": false,
          "title": "Active",
          "type": "boolean"
        },
        "base_seed": {
          "maximum": 2147483647,
          "minimum": 0,
          "title": "Base Seed",
          "type": "integer"
        },
        "benchmark_id": {
          "pattern": "^B-[0-9]{4}$",
          "title": "Benchmark Id",
          "type": "string"
        },
        "benchmark_version": {
          "const": 9,
          "default": 9,
          "title": "Benchmark Version",
          "type": "integer"
        },
        "cohort_id": {
          "pattern": "^[a-z0-9][a-z0-9-]{0,63}$",
          "title": "Cohort Id",
          "type": "string"
        },
        "display_name": {
          "maxLength": 160,
          "minLength": 1,
          "title": "Display Name",
          "type": "string"
        },
        "iterations": {
          "maximum": 100,
          "minimum": 1,
          "title": "Iterations",
          "type": "integer"
        },
        "max_questions": {
          "minimum": 1,
          "title": "Max Questions",
          "type": "integer"
        },
        "model_ids": {
          "items": {
            "type": "string"
          },
          "minItems": 1,
          "title": "Model Ids",
          "type": "array"
        },
        "target_ids": {
          "items": {
            "type": "string"
          },
          "minItems": 1,
          "title": "Target Ids",
          "type": "array"
        }
      },
      "required": [
        "cohort_id",
        "display_name",
        "benchmark_id",
        "target_ids",
        "iterations",
        "base_seed",
        "max_questions",
        "model_ids"
      ],
      "title": "CohortConfig",
      "type": "object"
    },
    "PublicActionTurn": {
      "additionalProperties": false,
      "properties": {
        "action": {
          "enum": [
            "ASK",
            "GUESS"
          ],
          "title": "Action",
          "type": "string"
        },
        "adjudicator": {
          "enum": [
            "oracle",
            "guess_validator"
          ],
          "title": "Adjudicator",
          "type": "string"
        },
        "answer": {
          "enum": [
            "YES",
            "NO",
            "UNKNOWN"
          ],
          "title": "Answer",
          "type": "string"
        },
        "counted": {
          "title": "Counted",
          "type": "boolean"
        },
        "counted_questions": {
          "title": "Counted Questions",
          "type": "integer"
        },
        "evidence": {
          "items": {
            "$ref": "#/$defs/PublicEvidence"
          },
          "title": "Evidence",
          "type": "array"
        },
        "guess_description": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Guess Description"
        },
        "guess_name": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Guess Name"
        },
        "oracle_cache": {
          "anyOf": [
            {
              "$ref": "#/$defs/PublicOracleCacheSource"
            },
            {
              "$ref": "#/$defs/PublicEpisodeOracleCacheSource"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Oracle Cache"
        },
        "question": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Question"
        },
        "recorded_output": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Recorded Output"
        },
        "turn_number": {
          "title": "Turn Number",
          "type": "integer"
        },
        "turn_type": {
          "const": "action",
          "default": "action",
          "title": "Turn Type",
          "type": "string"
        },
        "validator_explanation": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Validator Explanation"
        }
      },
      "required": [
        "turn_number",
        "action",
        "question",
        "guess_name",
        "guess_description",
        "adjudicator",
        "answer",
        "validator_explanation",
        "counted",
        "counted_questions",
        "evidence",
        "recorded_output"
      ],
      "title": "PublicActionTurn",
      "type": "object"
    },
    "PublicEpisodeDetail": {
      "additionalProperties": false,
      "properties": {
        "ask_count": {
          "title": "Ask Count",
          "type": "integer"
        },
        "cache_status": {
          "title": "Cache Status",
          "type": "string"
        },
        "completed_at": {
          "format": "date-time",
          "title": "Completed At",
          "type": "string"
        },
        "contract": {
          "$ref": "#/$defs/ContractReliabilitySnapshot"
        },
        "counted_questions": {
          "title": "Counted Questions",
          "type": "integer"
        },
        "duration_ms": {
          "title": "Duration Ms",
          "type": "integer"
        },
        "episode_id": {
          "title": "Episode Id",
          "type": "string"
        },
        "episode_run_id": {
          "title": "Episode Run Id",
          "type": "string"
        },
        "guess_count": {
          "title": "Guess Count",
          "type": "integer"
        },
        "guesser_disclosure": {
          "anyOf": [
            {
              "$ref": "#/$defs/PublicGuesserDisclosure"
            },
            {
              "type": "null"
            }
          ]
        },
        "oracle_unknown_count": {
          "title": "Oracle Unknown Count",
          "type": "integer"
        },
        "publication_eligible": {
          "title": "Publication Eligible",
          "type": "boolean"
        },
        "rejected_guess_count": {
          "title": "Rejected Guess Count",
          "type": "integer"
        },
        "scoring_eligible": {
          "title": "Scoring Eligible",
          "type": "boolean"
        },
        "started_at": {
          "format": "date-time",
          "title": "Started At",
          "type": "string"
        },
        "subject_description": {
          "title": "Subject Description",
          "type": "string"
        },
        "subject_name": {
          "title": "Subject Name",
          "type": "string"
        },
        "subject_reference_url": {
          "anyOf": [
            {
              "format": "uri",
              "maxLength": 2083,
              "minLength": 1,
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Subject Reference Url"
        },
        "success": {
          "title": "Success",
          "type": "boolean"
        },
        "telemetry": {
          "$ref": "#/$defs/PublicEpisodeTelemetry"
        },
        "terminal_reason": {
          "title": "Terminal Reason",
          "type": "string"
        },
        "total_cost_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Total Cost Usd",
          "type": "string"
        },
        "total_tokens": {
          "title": "Total Tokens",
          "type": "integer"
        },
        "total_turns": {
          "title": "Total Turns",
          "type": "integer"
        },
        "turns": {
          "items": {
            "discriminator": {
              "mapping": {
                "action": "#/$defs/PublicActionTurn",
                "contract_violation": "#/$defs/PublicContractViolationTurn"
              },
              "propertyName": "turn_type"
            },
            "oneOf": [
              {
                "$ref": "#/$defs/PublicActionTurn"
              },
              {
                "$ref": "#/$defs/PublicContractViolationTurn"
              }
            ]
          },
          "title": "Turns",
          "type": "array"
        }
      },
      "required": [
        "episode_run_id",
        "episode_id",
        "subject_name",
        "subject_description",
        "subject_reference_url",
        "started_at",
        "completed_at",
        "duration_ms",
        "success",
        "terminal_reason",
        "scoring_eligible",
        "publication_eligible",
        "total_turns",
        "counted_questions",
        "ask_count",
        "guess_count",
        "rejected_guess_count",
        "oracle_unknown_count",
        "cache_status",
        "total_cost_usd",
        "total_tokens",
        "contract",
        "guesser_disclosure",
        "telemetry",
        "turns"
      ],
      "title": "PublicEpisodeDetail",
      "type": "object"
    },
    "PublicRun": {
      "additionalProperties": false,
      "properties": {
        "base_seed": {
          "title": "Base Seed",
          "type": "integer"
        },
        "benchmark_id": {
          "title": "Benchmark Id",
          "type": "string"
        },
        "benchmark_mode": {
          "title": "Benchmark Mode",
          "type": "string"
        },
        "benchmark_name": {
          "title": "Benchmark Name",
          "type": "string"
        },
        "classification": {
          "enum": [
            "official",
            "lab"
          ],
          "title": "Classification",
          "type": "string"
        },
        "comparison": {
          "$ref": "#/$defs/PublicRunComparison"
        },
        "completed_at": {
          "format": "date-time",
          "title": "Completed At",
          "type": "string"
        },
        "contract": {
          "$ref": "#/$defs/ContractReliabilitySnapshot"
        },
        "created_at": {
          "format": "date-time",
          "title": "Created At",
          "type": "string"
        },
        "execution_id": {
          "title": "Execution Id",
          "type": "string"
        },
        "git_commit": {
          "title": "Git Commit",
          "type": "string"
        },
        "infrastructure_failed": {
          "title": "Infrastructure Failed",
          "type": "integer"
        },
        "iterations": {
          "title": "Iterations",
          "type": "integer"
        },
        "max_questions": {
          "title": "Max Questions",
          "type": "integer"
        },
        "model_failed": {
          "title": "Model Failed",
          "type": "integer"
        },
        "model_id": {
          "title": "Model Id",
          "type": "string"
        },
        "model_name": {
          "title": "Model Name",
          "type": "string"
        },
        "models": {
          "items": {
            "$ref": "#/$defs/PublicRunModel"
          },
          "title": "Models",
          "type": "array"
        },
        "question_score": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Question Score"
        },
        "question_score_confidence_interval": {
          "anyOf": [
            {
              "$ref": "#/$defs/QuestionScoreConfidenceInterval"
            },
            {
              "type": "null"
            }
          ],
          "default": null
        },
        "reason_codes": {
          "items": {
            "type": "string"
          },
          "title": "Reason Codes",
          "type": "array"
        },
        "subjects": {
          "items": {
            "$ref": "#/$defs/PublicSubject"
          },
          "title": "Subjects",
          "type": "array"
        },
        "success_rate": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Success Rate"
        },
        "successful": {
          "title": "Successful",
          "type": "integer"
        },
        "target_ids": {
          "items": {
            "type": "string"
          },
          "title": "Target Ids",
          "type": "array"
        },
        "terminal_trials": {
          "title": "Terminal Trials",
          "type": "integer"
        },
        "total_cost_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Total Cost Usd",
          "type": "string"
        },
        "totals": {
          "$ref": "#/$defs/PublicRunTotals"
        }
      },
      "required": [
        "execution_id",
        "model_id",
        "model_name",
        "benchmark_id",
        "benchmark_name",
        "classification",
        "reason_codes",
        "completed_at",
        "created_at",
        "git_commit",
        "benchmark_mode",
        "target_ids",
        "iterations",
        "base_seed",
        "max_questions",
        "success_rate",
        "question_score",
        "total_cost_usd",
        "successful",
        "model_failed",
        "infrastructure_failed",
        "terminal_trials",
        "contract",
        "totals",
        "comparison",
        "models",
        "subjects"
      ],
      "title": "PublicRun",
      "type": "object"
    },
    "PublicSubject": {
      "additionalProperties": false,
      "properties": {
        "average_questions": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Average Questions"
        },
        "contract": {
          "$ref": "#/$defs/ContractReliabilitySnapshot"
        },
        "display_name": {
          "title": "Display Name",
          "type": "string"
        },
        "entity_type": {
          "title": "Entity Type",
          "type": "string"
        },
        "infrastructure_failed": {
          "title": "Infrastructure Failed",
          "type": "integer"
        },
        "model_failed": {
          "title": "Model Failed",
          "type": "integer"
        },
        "success_rate": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Success Rate"
        },
        "successful": {
          "title": "Successful",
          "type": "integer"
        },
        "target_id": {
          "title": "Target Id",
          "type": "string"
        },
        "trials": {
          "items": {
            "$ref": "#/$defs/PublicTrial"
          },
          "title": "Trials",
          "type": "array"
        }
      },
      "required": [
        "target_id",
        "display_name",
        "entity_type",
        "success_rate",
        "average_questions",
        "successful",
        "model_failed",
        "infrastructure_failed",
        "contract",
        "trials"
      ],
      "title": "PublicSubject",
      "type": "object"
    },
    "PublicTrial": {
      "additionalProperties": false,
      "properties": {
        "contract": {
          "anyOf": [
            {
              "$ref": "#/$defs/ContractReliabilitySnapshot"
            },
            {
              "type": "null"
            }
          ]
        },
        "cost_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Cost Usd",
          "type": "string"
        },
        "counted_questions": {
          "title": "Counted Questions",
          "type": "integer"
        },
        "duration_ms": {
          "title": "Duration Ms",
          "type": "integer"
        },
        "episode": {
          "anyOf": [
            {
              "$ref": "#/$defs/PublicEpisodeDetail"
            },
            {
              "type": "null"
            }
          ],
          "default": null
        },
        "failure_code": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Failure Code"
        },
        "penalized_questions": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Penalized Questions"
        },
        "status": {
          "enum": [
            "success",
            "model_failure",
            "infrastructure_failure"
          ],
          "title": "Status",
          "type": "string"
        },
        "trial_id": {
          "title": "Trial Id",
          "type": "string"
        },
        "trial_number": {
          "title": "Trial Number",
          "type": "integer"
        }
      },
      "required": [
        "trial_id",
        "trial_number",
        "status",
        "counted_questions",
        "penalized_questions",
        "cost_usd",
        "duration_ms",
        "contract"
      ],
      "title": "PublicTrial",
      "type": "object"
    },
    "PublicComponentTelemetry": {
      "additionalProperties": false,
      "properties": {
        "cache_write_tokens": {
          "title": "Cache Write Tokens",
          "type": "integer"
        },
        "cached_input_tokens": {
          "title": "Cached Input Tokens",
          "type": "integer"
        },
        "calls": {
          "title": "Calls",
          "type": "integer"
        },
        "cost_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Cost Usd",
          "type": "string"
        },
        "estimated_cache_savings_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Estimated Cache Savings Usd",
          "type": "string"
        },
        "input_tokens": {
          "title": "Input Tokens",
          "type": "integer"
        },
        "latency_ms": {
          "title": "Latency Ms",
          "type": "integer"
        },
        "output_tokens": {
          "title": "Output Tokens",
          "type": "integer"
        },
        "reasoning_tokens": {
          "title": "Reasoning Tokens",
          "type": "integer"
        },
        "total_tokens": {
          "title": "Total Tokens",
          "type": "integer"
        }
      },
      "required": [
        "calls",
        "cost_usd",
        "latency_ms",
        "total_tokens",
        "input_tokens",
        "cached_input_tokens",
        "cache_write_tokens",
        "output_tokens",
        "reasoning_tokens",
        "estimated_cache_savings_usd"
      ],
      "title": "PublicComponentTelemetry",
      "type": "object"
    },
    "PublicContractViolationTurn": {
      "additionalProperties": false,
      "properties": {
        "counted": {
          "title": "Counted",
          "type": "boolean"
        },
        "counted_questions": {
          "title": "Counted Questions",
          "type": "integer"
        },
        "feedback_event": {
          "anyOf": [
            {
              "const": "FORMAT_ERROR",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Feedback Event"
        },
        "rejected_outputs": {
          "default": [],
          "items": {
            "$ref": "#/$defs/PublicRejectedOutput"
          },
          "title": "Rejected Outputs",
          "type": "array"
        },
        "turn_number": {
          "title": "Turn Number",
          "type": "integer"
        },
        "turn_type": {
          "const": "contract_violation",
          "default": "contract_violation",
          "title": "Turn Type",
          "type": "string"
        },
        "violation_code": {
          "const": "invalid_guesser_output",
          "title": "Violation Code",
          "type": "string"
        },
        "violation_kind": {
          "enum": [
            "invalid_json",
            "invalid_action",
            "output_limit_exceeded",
            "empty_output",
            "incomplete_output"
          ],
          "title": "Violation Kind",
          "type": "string"
        }
      },
      "required": [
        "turn_number",
        "violation_code",
        "violation_kind",
        "feedback_event",
        "counted",
        "counted_questions"
      ],
      "title": "PublicContractViolationTurn",
      "type": "object"
    },
    "PublicEpisodeOracleCacheSource": {
      "additionalProperties": false,
      "properties": {
        "answered_at": {
          "format": "date-time",
          "title": "Answered At",
          "type": "string"
        },
        "episode_id": {
          "title": "Episode Id",
          "type": "string"
        },
        "question": {
          "title": "Question",
          "type": "string"
        },
        "scope": {
          "const": "same_episode",
          "default": "same_episode",
          "title": "Scope",
          "type": "string"
        },
        "target_id": {
          "title": "Target Id",
          "type": "string"
        },
        "turn_number": {
          "minimum": 1,
          "title": "Turn Number",
          "type": "integer"
        }
      },
      "required": [
        "target_id",
        "episode_id",
        "turn_number",
        "question",
        "answered_at"
      ],
      "title": "PublicEpisodeOracleCacheSource",
      "type": "object"
    },
    "PublicEpisodeTelemetry": {
      "additionalProperties": false,
      "properties": {
        "guesser": {
          "$ref": "#/$defs/PublicComponentTelemetry"
        },
        "oracle": {
          "$ref": "#/$defs/PublicComponentTelemetry"
        },
        "validator": {
          "$ref": "#/$defs/PublicComponentTelemetry"
        }
      },
      "required": [
        "guesser",
        "oracle",
        "validator"
      ],
      "title": "PublicEpisodeTelemetry",
      "type": "object"
    },
    "PublicEvidence": {
      "additionalProperties": false,
      "properties": {
        "excerpt": {
          "title": "Excerpt",
          "type": "string"
        },
        "kind": {
          "default": "quotation",
          "enum": [
            "quotation",
            "source_summary"
          ],
          "title": "Kind",
          "type": "string"
        },
        "source_url": {
          "format": "uri",
          "maxLength": 2083,
          "minLength": 1,
          "title": "Source Url",
          "type": "string"
        },
        "validation": {
          "const": "model_reported",
          "title": "Validation",
          "type": "string"
        }
      },
      "required": [
        "source_url",
        "excerpt",
        "validation"
      ],
      "title": "PublicEvidence",
      "type": "object"
    },
    "PublicExcludedRepairCost": {
      "additionalProperties": false,
      "properties": {
        "affected_trials": {
          "default": 0,
          "minimum": 0,
          "title": "Affected Trials",
          "type": "integer"
        },
        "cost_usd": {
          "default": "0",
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Cost Usd",
          "type": "string"
        },
        "superseded_attempts": {
          "default": 0,
          "minimum": 0,
          "title": "Superseded Attempts",
          "type": "integer"
        }
      },
      "title": "PublicExcludedRepairCost",
      "type": "object"
    },
    "PublicGuesserDisclosure": {
      "additionalProperties": false,
      "properties": {
        "begin_message": {
          "title": "Begin Message",
          "type": "string"
        },
        "output_storage": {
          "const": "canonical_structured_action",
          "default": "canonical_structured_action",
          "title": "Output Storage",
          "type": "string"
        },
        "required_formats": {
          "anyOf": [
            {
              "$ref": "#/$defs/PublicGuesserRequiredFormats"
            },
            {
              "type": "null"
            }
          ],
          "default": null
        },
        "system_message": {
          "title": "System Message",
          "type": "string"
        }
      },
      "required": [
        "system_message",
        "begin_message"
      ],
      "title": "PublicGuesserDisclosure",
      "type": "object"
    },
    "PublicGuesserRequiredFormats": {
      "additionalProperties": false,
      "properties": {
        "ask": {
          "minLength": 1,
          "title": "Ask",
          "type": "string"
        },
        "guess": {
          "minLength": 1,
          "title": "Guess",
          "type": "string"
        }
      },
      "required": [
        "ask",
        "guess"
      ],
      "title": "PublicGuesserRequiredFormats",
      "type": "object"
    },
    "PublicModel": {
      "additionalProperties": false,
      "properties": {
        "configuration_hash": {
          "title": "Configuration Hash",
          "type": "string"
        },
        "display_name": {
          "title": "Display Name",
          "type": "string"
        },
        "model_id": {
          "title": "Model Id",
          "type": "string"
        },
        "provider": {
          "title": "Provider",
          "type": "string"
        },
        "reasoning_effort": {
          "title": "Reasoning Effort",
          "type": "string"
        },
        "route": {
          "title": "Route",
          "type": "string"
        },
        "seed_capability": {
          "title": "Seed Capability",
          "type": "string"
        }
      },
      "required": [
        "model_id",
        "display_name",
        "route",
        "provider",
        "reasoning_effort",
        "seed_capability",
        "configuration_hash"
      ],
      "title": "PublicModel",
      "type": "object"
    },
    "PublicOracleCacheSource": {
      "additionalProperties": false,
      "description": "Allowlisted source attribution, without internal paths or provider call identifiers.",
      "properties": {
        "answered_at": {
          "format": "date-time",
          "title": "Answered At",
          "type": "string"
        },
        "benchmark_id": {
          "title": "Benchmark Id",
          "type": "string"
        },
        "episode_id": {
          "title": "Episode Id",
          "type": "string"
        },
        "execution_id": {
          "title": "Execution Id",
          "type": "string"
        },
        "model_id": {
          "title": "Model Id",
          "type": "string"
        },
        "question": {
          "title": "Question",
          "type": "string"
        },
        "scope": {
          "default": "historical",
          "enum": [
            "historical",
            "same_execution"
          ],
          "title": "Scope",
          "type": "string"
        },
        "target_id": {
          "title": "Target Id",
          "type": "string"
        },
        "trial_id": {
          "title": "Trial Id",
          "type": "string"
        },
        "turn_number": {
          "minimum": 1,
          "title": "Turn Number",
          "type": "integer"
        }
      },
      "required": [
        "execution_id",
        "model_id",
        "benchmark_id",
        "target_id",
        "trial_id",
        "episode_id",
        "turn_number",
        "question",
        "answered_at"
      ],
      "title": "PublicOracleCacheSource",
      "type": "object"
    },
    "PublicRejectedOutput": {
      "additionalProperties": false,
      "description": "Exact Guesser-visible provider text with private provider IDs removed.",
      "properties": {
        "attempt_number": {
          "minimum": 1,
          "title": "Attempt Number",
          "type": "integer"
        },
        "finish_reason": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Finish Reason"
        },
        "text": {
          "minLength": 1,
          "title": "Text",
          "type": "string"
        }
      },
      "required": [
        "attempt_number",
        "text"
      ],
      "title": "PublicRejectedOutput",
      "type": "object"
    },
    "PublicRunComparison": {
      "additionalProperties": false,
      "properties": {
        "cost_adjusted_question_score": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Cost Adjusted Question Score"
        },
        "efficiency_status": {
          "enum": [
            "ranked",
            "question_score_unavailable",
            "recorded_guesser_cost_unavailable",
            "no_terminal_episodes",
            "no_guesser_calls"
          ],
          "title": "Efficiency Status",
          "type": "string"
        },
        "full_cost_per_episode_usd": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Full Cost Per Episode Usd"
        },
        "guesser_cost_per_episode_usd": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Guesser Cost Per Episode Usd"
        },
        "guesser_latency_per_call_ms": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Guesser Latency Per Call Ms"
        },
        "guesser_think_time_per_episode_ms": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Guesser Think Time Per Episode Ms"
        },
        "runtime_per_episode_ms": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Runtime Per Episode Ms"
        },
        "support_cost_per_episode_usd": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Support Cost Per Episode Usd"
        },
        "support_cost_share": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Support Cost Share"
        }
      },
      "required": [
        "efficiency_status"
      ],
      "title": "PublicRunComparison",
      "type": "object"
    },
    "PublicRunCostTotals": {
      "additionalProperties": false,
      "properties": {
        "guesser": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Guesser",
          "type": "string"
        },
        "judge": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Judge",
          "type": "string"
        },
        "primary_oracle": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Primary Oracle",
          "type": "string"
        },
        "reviewer": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Reviewer",
          "type": "string"
        },
        "total": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Total",
          "type": "string"
        },
        "validator": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Validator",
          "type": "string"
        }
      },
      "required": [
        "guesser",
        "primary_oracle",
        "reviewer",
        "judge",
        "validator",
        "total"
      ],
      "title": "PublicRunCostTotals",
      "type": "object"
    },
    "PublicRunModel": {
      "additionalProperties": false,
      "properties": {
        "calls": {
          "minimum": 0,
          "title": "Calls",
          "type": "integer"
        },
        "configuration_id": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Configuration Id"
        },
        "cost_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Cost Usd",
          "type": "string"
        },
        "fallback_calls": {
          "default": 0,
          "minimum": 0,
          "title": "Fallback Calls",
          "type": "integer"
        },
        "prompt_version": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "title": "Prompt Version"
        },
        "prompt_versions": {
          "default": [],
          "items": {
            "type": "string"
          },
          "title": "Prompt Versions",
          "type": "array"
        },
        "provider_routing": {
          "default": "exact",
          "enum": [
            "exact",
            "automatic"
          ],
          "title": "Provider Routing",
          "type": "string"
        },
        "providers": {
          "default": [],
          "items": {
            "$ref": "#/$defs/ResolvedProviderUsageSnapshot"
          },
          "title": "Providers",
          "type": "array"
        },
        "reasoning_effort": {
          "title": "Reasoning Effort",
          "type": "string"
        },
        "requested_model": {
          "title": "Requested Model",
          "type": "string"
        },
        "requested_provider": {
          "title": "Requested Provider",
          "type": "string"
        },
        "resolved_models": {
          "items": {
            "type": "string"
          },
          "title": "Resolved Models",
          "type": "array"
        },
        "resolved_providers": {
          "items": {
            "type": "string"
          },
          "title": "Resolved Providers",
          "type": "array"
        },
        "role": {
          "enum": [
            "guesser",
            "oracle",
            "reviewer",
            "judge",
            "validator"
          ],
          "title": "Role",
          "type": "string"
        },
        "unreported_calls": {
          "default": 0,
          "minimum": 0,
          "title": "Unreported Calls",
          "type": "integer"
        }
      },
      "required": [
        "role",
        "configuration_id",
        "requested_model",
        "requested_provider",
        "resolved_models",
        "resolved_providers",
        "reasoning_effort",
        "prompt_version",
        "calls",
        "cost_usd"
      ],
      "title": "PublicRunModel",
      "type": "object"
    },
    "PublicRunTotals": {
      "additionalProperties": false,
      "properties": {
        "costs_usd": {
          "$ref": "#/$defs/PublicRunCostTotals"
        },
        "excluded_repair": {
          "$ref": "#/$defs/PublicExcludedRepairCost"
        },
        "guesser_calls": {
          "minimum": 0,
          "title": "Guesser Calls",
          "type": "integer"
        },
        "guesser_think_time_ms": {
          "minimum": 0,
          "title": "Guesser Think Time Ms",
          "type": "integer"
        },
        "runtime_ms": {
          "minimum": 0,
          "title": "Runtime Ms",
          "type": "integer"
        },
        "total_tokens": {
          "minimum": 0,
          "title": "Total Tokens",
          "type": "integer"
        }
      },
      "required": [
        "costs_usd",
        "total_tokens",
        "runtime_ms",
        "guesser_think_time_ms",
        "guesser_calls"
      ],
      "title": "PublicRunTotals",
      "type": "object"
    },
    "QuestionScoreConfidenceInterval": {
      "additionalProperties": false,
      "properties": {
        "confidence_level": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Confidence Level",
          "type": "string"
        },
        "degrees_of_freedom": {
          "anyOf": [
            {
              "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "title": "Degrees Of Freedom"
        },
        "estimate": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Estimate",
          "type": "string"
        },
        "lower": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Lower",
          "type": "string"
        },
        "method": {
          "const": "stratified-welch-t-v1",
          "title": "Method",
          "type": "string"
        },
        "standard_error": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Standard Error",
          "type": "string"
        },
        "subject_count": {
          "minimum": 1,
          "title": "Subject Count",
          "type": "integer"
        },
        "trial_count": {
          "minimum": 2,
          "title": "Trial Count",
          "type": "integer"
        },
        "upper": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Upper",
          "type": "string"
        }
      },
      "required": [
        "confidence_level",
        "method",
        "estimate",
        "lower",
        "upper",
        "standard_error",
        "subject_count",
        "trial_count"
      ],
      "title": "QuestionScoreConfidenceInterval",
      "type": "object"
    },
    "ResolvedProviderUsageSnapshot": {
      "additionalProperties": false,
      "properties": {
        "calls": {
          "minimum": 1,
          "title": "Calls",
          "type": "integer"
        },
        "cost_usd": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Cost Usd",
          "type": "string"
        },
        "latency_ms": {
          "minimum": 0,
          "title": "Latency Ms",
          "type": "integer"
        },
        "provider": {
          "minLength": 1,
          "title": "Provider",
          "type": "string"
        }
      },
      "required": [
        "provider",
        "calls",
        "cost_usd",
        "latency_ms"
      ],
      "title": "ResolvedProviderUsageSnapshot",
      "type": "object"
    },
    "ScorePolicy": {
      "additionalProperties": false,
      "properties": {
        "failure_penalty_offset": {
          "default": 1,
          "maximum": 100,
          "minimum": 1,
          "title": "Failure Penalty Offset",
          "type": "integer"
        },
        "version": {
          "const": "average-then-average-v1",
          "default": "average-then-average-v1",
          "title": "Version",
          "type": "string"
        }
      },
      "title": "ScorePolicy",
      "type": "object"
    },
    "SiteMetadata": {
      "additionalProperties": false,
      "properties": {
        "base_path": {
          "pattern": "^/(?:[A-Za-z0-9._/-]*/)?$",
          "title": "Base Path",
          "type": "string"
        },
        "citation_label": {
          "maxLength": 240,
          "minLength": 1,
          "title": "Citation Label",
          "type": "string"
        },
        "creator_name": {
          "maxLength": 160,
          "minLength": 1,
          "title": "Creator Name",
          "type": "string"
        },
        "description": {
          "maxLength": 300,
          "minLength": 1,
          "title": "Description",
          "type": "string"
        },
        "short_title": {
          "maxLength": 40,
          "minLength": 1,
          "title": "Short Title",
          "type": "string"
        },
        "title": {
          "maxLength": 100,
          "minLength": 1,
          "title": "Title",
          "type": "string"
        }
      },
      "required": [
        "title",
        "short_title",
        "description",
        "base_path",
        "creator_name",
        "citation_label"
      ],
      "title": "SiteMetadata",
      "type": "object"
    },
    "Winner": {
      "additionalProperties": false,
      "properties": {
        "display_names": {
          "items": {
            "type": "string"
          },
          "title": "Display Names",
          "type": "array"
        },
        "joint": {
          "title": "Joint",
          "type": "boolean"
        },
        "model_ids": {
          "items": {
            "type": "string"
          },
          "title": "Model Ids",
          "type": "array"
        },
        "question_score": {
          "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d*$",
          "title": "Question Score",
          "type": "string"
        }
      },
      "required": [
        "model_ids",
        "display_names",
        "question_score",
        "joint"
      ],
      "title": "Winner",
      "type": "object"
    }
  },
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "additionalProperties": false,
  "properties": {
    "active_cohort": {
      "$ref": "#/$defs/CohortConfig"
    },
    "lab_runs": {
      "items": {
        "$ref": "#/$defs/PublicRun"
      },
      "title": "Lab Runs",
      "type": "array"
    },
    "leaderboard": {
      "items": {
        "$ref": "#/$defs/LeaderboardRow"
      },
      "title": "Leaderboard",
      "type": "array"
    },
    "models": {
      "items": {
        "$ref": "#/$defs/PublicModel"
      },
      "title": "Models",
      "type": "array"
    },
    "official_runs": {
      "items": {
        "$ref": "#/$defs/PublicRun"
      },
      "title": "Official Runs",
      "type": "array"
    },
    "provenance": {
      "$ref": "#/$defs/DatasetProvenance"
    },
    "schema_version": {
      "const": 9,
      "default": 9,
      "title": "Schema Version",
      "type": "integer"
    },
    "score_policy": {
      "$ref": "#/$defs/ScorePolicy"
    },
    "site": {
      "$ref": "#/$defs/SiteMetadata"
    },
    "winner": {
      "anyOf": [
        {
          "$ref": "#/$defs/Winner"
        },
        {
          "type": "null"
        }
      ]
    }
  },
  "required": [
    "site",
    "score_policy",
    "active_cohort",
    "provenance",
    "winner",
    "leaderboard",
    "models",
    "official_runs",
    "lab_runs"
  ],
  "title": "PublishedDataset",
  "type": "object"
}