{
  "authorization": {
    "base_campaign_authorized": false,
    "fresh_pilot_authorized": false,
    "fresh_pilot_completed": true,
    "fresh_pilot_decision": "sealed-pilot-quarantined-provider-failure",
    "fresh_pilot_expected_cost_usd": "15.00",
    "fresh_pilot_proposed_cap_usd": "25.00",
    "fresh_pilot_proposed_provider_caps_usd": {
      "Anthropic": "15.00",
      "OpenAI": "10.00"
    },
    "fresh_pilot_v2_authorized": true,
    "fresh_pilot_v2_completed": true,
    "fresh_pilot_v2_decision": "fresh-pilot-v2-quarantined-engineering-only",
    "fresh_pilot_v2_expected_cost_usd": "11.51",
    "fresh_pilot_v2_hard_cap_usd": "25",
    "fresh_pilot_v3_authorized": true,
    "fresh_pilot_v3_completed": true,
    "fresh_pilot_v3_cost_usd": "13.1861145",
    "fresh_pilot_v3_decision": "fresh-pilot-v3-quarantined-engineering-only",
    "fresh_pilot_v3_hard_cap_usd": "25",
    "further_provider_calls_authorized": false,
    "original_pilot_cap_usd": 100,
    "public_calibration_cap_usd": 20,
    "public_calibration_cost_usd": "2.311758000",
    "public_calibration_was_authorized": true,
    "sealed_pilot_completed": true
  },
  "campaign_decision": "sealed-pilot-quarantined-provider-failure",
  "display": {
    "formal_suite": "TreasureBench",
    "historical_alias": "DiscoveryBench",
    "role": "formal-instrument"
  },
  "encrypted_private_material_retained": true,
  "expected_future_study_id": "DD-023",
  "future_base": {
    "batches": 4,
    "boundary_repeats": 62,
    "canonical_cells": 138,
    "instances": 200,
    "repeats": 3,
    "status": "blocked-local-open-host-ineligible"
  },
  "instrument_id": "discoverybench-agents-v1",
  "instrument_owner": "DD-010",
  "models": {
    "cloud_snapshots": [
      "gpt-5.4-2026-03-05",
      "claude-sonnet-4-6"
    ],
    "downloads": 0,
    "local_revision": "mistralai/Mistral-Small-3.1-24B-Instruct-2503@68faf511d618ef198fef186659617cfd2eb8e33a",
    "public_calibration_calls_cumulative": 607
  },
  "next_gate": "no replacement private pilot selected; explicit owner decision required; base campaign blocked",
  "private_material_published": false,
  "provider_comparative_results_published": false,
  "provider_schema_conformance": {
    "bisection_calls": 0,
    "calls": 4,
    "canonical_exactly_one_final_action": "pass",
    "classification": "public-engineering-only-not-scientific-evidence",
    "cost_usd": "0.0086685",
    "decision": "conformance-pass-both-complete-schemas",
    "execution_commit": "dcd89bca5daa54386df163439a010583c30134ec",
    "input_tokens": 1603,
    "output_tokens": 275,
    "performance_evidence": false,
    "private_state_changed": false,
    "provider_cost_usd": {
      "anthropic": "0.005481",
      "openai": "0.0031875"
    },
    "scientific_state_changed": false
  },
  "public_calibration": {
    "architectures": 5,
    "classification": "engineering-only-not-scientific-evidence",
    "composite_score": null,
    "contamination_clear": true,
    "cumulative_calls_across_attempts": 607,
    "cumulative_cost_usd": "2.311758000",
    "method_a_b_agree": true,
    "private_material_exists": false,
    "protocol_compliance": true,
    "provider_ranking": false,
    "required_direct_routes": 2,
    "route_cases": 100,
    "scientific_evidence_exists": false,
    "tasks": 10
  },
  "repair_adjudication": {
    "decision": "instrument-repaired-fresh-sealed-pilot-required",
    "dimension_breakdowns_published": false,
    "invalid_final_agent_outputs": 266,
    "legacy_coverage_range_violations": 57,
    "metric_records_changed_by_extra_action_credit": 138,
    "over_budget_final_agent_outputs": 265,
    "performance_results_published": false,
    "private_run_traces_inspected": 500,
    "protocol_invalid_runs_attributable_to_terminal_provider_missingness": 1,
    "provider_events": [
      {
        "coarse_error_class": "transient-provider",
        "provider": "Anthropic",
        "status": "recovered",
        "taxonomy_class": "provider-service-recovered",
        "terminal_run_missing": false
      },
      {
        "coarse_error_class": "schema-or-parameter",
        "provider": "Anthropic",
        "status": "terminal",
        "taxonomy_class": "unknown-terminal",
        "terminal_run_missing": true
      }
    ],
    "runs_with_invalid_final_cardinality": 137,
    "separate_downstream_protocol_invalid_runs": 1,
    "status": "complete-redacted-engineering-adjudication"
  },
  "repair_confirmation_pilot": {
    "answer_ciphertext_created": false,
    "answer_key_created": false,
    "answers_created": 0,
    "anthropic_calls": 0,
    "attempts": 1,
    "authorization_inactive_archived": true,
    "batch_id": "tb-agents-v1-repair-confirmation-v1-b01",
    "campaign_id": "treasurebench-agents-v1-repair-confirmation-v1",
    "cost_usd": "0.00",
    "custody_manifest_created": false,
    "decision": "sealed-pilot-quarantined-provider-failure",
    "error_class": "schema-or-parameter",
    "errors": 1,
    "execution_commit": "fe313602df7f4e8ffac1a1a02c2b3a83f3c72943",
    "input_tokens": 0,
    "name": "fresh-repair-confirmation-pilot",
    "openai_calls": 1,
    "output_lock_commitment": "sha256:8102a6c1b6bda003336d5503136dfe29301b04cb8f35e7740edd8d56f0eb3c1d",
    "output_tokens": 0,
    "performance_results_published": false,
    "private_runs": 0,
    "provider_phase_closed": true,
    "stage": "public-canary",
    "status": "quarantined",
    "successes": 0,
    "task_ciphertext_created": false,
    "task_key_created": false,
    "task_seed_created": false,
    "tasks_created": 0,
    "terminal_http_status": 400
  },
  "repair_confirmation_pilot_v2": {
    "answer_ciphertext_created": false,
    "answer_key_created": true,
    "anthropic_calls": 1,
    "attempts": 2,
    "batch_id": "tb-agents-v1-repair-confirmation-v2-b01",
    "campaign_id": "treasurebench-agents-v1-repair-confirmation-v2",
    "cost_usd": "0.0076095",
    "custody_manifest_created": false,
    "decision": "fresh-pilot-v2-quarantined-engineering-only",
    "errors": 0,
    "execution_commit": "d210b0653677859c79a1fb87d871aaf45f4a81d4",
    "failure_class": "custody-creation-failure",
    "input_tokens": 1349,
    "name": "fresh-repair-confirmation-pilot-v2",
    "objects_locked": 8,
    "openai_calls": 1,
    "output_lock_commitment": "sha256:127a9c796459c7627f6fd90b92ef1587ad0f6b1910b4ff255c2ceb976f3ab25f",
    "output_tokens": 253,
    "performance_results_published": false,
    "private_runs": 0,
    "provider_cost_usd": {
      "Anthropic": "0.004812",
      "OpenAI": "0.0027975"
    },
    "provider_phase_closed": true,
    "stage": "custody",
    "status": "quarantined",
    "successes": 2,
    "task_ciphertext_created": false,
    "task_key_created": true,
    "task_seed_created": true,
    "unsealed": false
  },
  "repair_confirmation_pilot_v3": {
    "batch_id": "tb-agents-v1-repair-confirmation-v3-b01",
    "calls": 3067,
    "campaign_id": "treasurebench-agents-v1-repair-confirmation-v3",
    "cost_usd": "13.1861145",
    "custody_complete": true,
    "decision": "fresh-pilot-v3-quarantined-engineering-only",
    "execution_commit": "0f9d82bb50cbb334bea47e24448831faf0cdbed8",
    "failure_class": "fixed-full-batch-failure",
    "fixed_full_batch_complete": false,
    "input_tokens": 2304303,
    "name": "fresh-repair-confirmation-pilot-v3",
    "objects_locked": 3576,
    "output_lock_commitment": "sha256:e52055b08ca3a8acb1cfb6ac608c6e601f3c618352900f92bf91c5ffc4718dbb",
    "output_tokens": 444085,
    "partial_private_pairing_count": "not-published-quarantined-batch",
    "performance_results_published": false,
    "private_prefix_complete": true,
    "provider_cost_usd": {
      "Anthropic": "8.590857",
      "OpenAI": "4.5952575"
    },
    "provider_phase_closed": true,
    "public_canary_complete": true,
    "stage": "fixed-full-batch",
    "status": "quarantined",
    "unsealed": false
  },
  "schema_version": 1,
  "scientific_results_exist": false,
  "selected_next_tier": {
    "architectures": 5,
    "attempts": 3037,
    "batches": 1,
    "claim_eligible": false,
    "contamination_findings": 0,
    "cost_usd": "11.5702435",
    "decision": "sealed-pilot-quarantined-provider-failure",
    "errors": 2,
    "instances": 50,
    "method_a_b_disagreements": 0,
    "name": "sealed-engineering-pilot",
    "output_lock_commitment": "sha256:1d487723f7587e8e2fa865682e6f6cc473cf2da4967b837dedf3952cddfcbfab",
    "private_runs": 500,
    "protocol_errors": 2,
    "repeats": 1,
    "status": "complete-quarantined",
    "successes": 3035
  },
  "study_id_allocated": false
}