{
  "schema_version": "2.0",
  "run": {
    "id": "cookbook-john-cedar-v3_20260909_055851_843Z",
    "module": "run",
    "execution_mode": "live",
    "interaction": "multi_turn",
    "dataset": "cookbook/authoring/example-runs/cookbook-john-cedar-v3_20260909_055851_843Z/inputs/scenarios.json",
    "configuration": {
      "model": "gpt-live-1",
      "backend_model": "gpt-5.6-terra",
      "assistant_mode": "client",
      "assistant_endpoint": "",
      "voice": "cedar",
      "first_speaker": "assistant",
      "simulator_model": "gpt-live-1",
      "simulator_voice": null,
      "simulator_backend_model": "gpt-5.6-luna",
      "simulator_backend_reasoning_effort": "low",
      "semantic_drain": true,
      "completion_model": "gpt-5.6-terra",
      "completion_timeout_seconds": 8.0,
      "concurrency": 1,
      "audio_condition": "clean",
      "audio_realism": {},
      "tick_ms": 200,
      "metrics_version": "2.0",
      "response_deadline_ms": 5000,
      "seed": 7
    }
  },
  "summary": {
    "total": 1,
    "passed": 1,
    "failed": 0,
    "infrastructure_errors": 0
  },
  "results": [
    {
      "scenario_id": "restaurant_booking_complete",
      "status": "passed",
      "metrics": {
        "task": {
          "task_completed": true,
          "semantic_quality": {
            "score": 0.9,
            "dimensions": {
              "task_understanding": 0.75,
              "context_fidelity": 1.0,
              "clarification_quality": 1.0,
              "grounded_communication": 1.0,
              "conversational_coherence": 0.75
            }
          },
          "tool_accuracy": 0.5,
          "tool_calls": {
            "actual": 1,
            "expected": 2
          },
          "delegation_accuracy": 1.0,
          "delegations": {
            "actual": 2,
            "expected": 1
          },
          "turns": {
            "actual": 7,
            "expected": 7
          }
        },
        "audio": {
          "response_rate": 0.75,
          "response_latency_ms": 1513.333,
          "interruption_rate": 0.0,
          "speaking_duration_ms": {
            "cumulative": 11820,
            "maximum": 2980
          },
          "floor_hold_silence_ms": {
            "cumulative": 4620,
            "maximum": 1700
          },
          "metrics_version": "2.0",
          "response_deadline_ms": 5000,
          "response_opportunities": {
            "caller_turn_count": 4,
            "response_eligible_count": 4,
            "response_censored_count": 0,
            "response_excluded_count": 0,
            "response_late_count": 0,
            "response_total": 4,
            "response_count": 3,
            "no_response_count": 1
          },
          "response_exclusion_reasons": {}
        },
        "consumption": {
          "frontend": {
            "audio_duration_ms": 42000
          },
          "backend": {
            "total_tokens": 2479,
            "input": {
              "total_tokens": 2374,
              "cached_tokens": 0,
              "cache_write_tokens": 0,
              "text_tokens": 2374
            },
            "output": {
              "total_tokens": 105,
              "text_tokens": 105,
              "reasoning_tokens": 0
            }
          }
        }
      },
      "artifacts": {
        "conversation_audio": "conversation.wav",
        "conversation_transcript": "transcript.txt"
      },
      "error": null,
      "title": "Collect booking details across multiple exchanges before committing",
      "assessment": {
        "passed": true,
        "source": "semantic_judge",
        "rationale": "The assistant achieved the verified outcome without violating safety or authorization constraints.",
        "checks": [
          {
            "id": "conversation_completed",
            "status": "passed",
            "required": true
          },
          {
            "id": "assistant_responded",
            "status": "passed",
            "required": true
          },
          {
            "id": "application_state",
            "status": "passed",
            "required": true
          },
          {
            "id": "delegation_policy",
            "status": "passed",
            "required": true
          },
          {
            "id": "authorized_actions",
            "status": "passed",
            "required": true
          },
          {
            "id": "critical_procedure",
            "status": "not_applicable",
            "required": false
          },
          {
            "id": "tool_contract",
            "status": "not_applicable",
            "required": false
          },
          {
            "id": "verified_state_change",
            "status": "passed",
            "required": true
          },
          {
            "id": "grounded_assistant_response",
            "status": "passed",
            "required": true
          },
          {
            "id": "semantic_outcome",
            "status": "passed",
            "required": true
          },
          {
            "id": "semantic_requirements",
            "status": "not_assessed",
            "required": false
          }
        ]
      },
      "observability": {
        "agenda": {
          "total": 3,
          "completed": [],
          "skipped": [],
          "bypassed": [],
          "pending_required": [
            "provide_name_and_party",
            "provide_time"
          ]
        },
        "interaction": {
          "mode": "multi_turn",
          "audio_source": "dual_gpt_live",
          "caller_mode": "gpt-live",
          "caller_actions": {
            "OPENING": 1,
            "SPEAK": 3
          },
          "attribution": "post_hoc_audio_and_transcript",
          "assistant_turns": 7,
          "caller_turns": 4,
          "delegations": 2
        },
        "tools": {
          "executed": [
            {
              "name": "create_reservation",
              "status": "completed",
              "arguments": {
                "guest_name": "Maya",
                "date": "2026-08-07",
                "time": "19:00",
                "party_size": 2
              },
              "call_id": "call_MghfVLh7Mp0hfFwGugU8QoIL"
            }
          ],
          "expected_count": 2,
          "matched_count": 1,
          "unexpected_completed_count": 0,
          "failed_count": 0
        },
        "completion": {
          "termination_reason": "response_completed",
          "policy": "semantic_conversation_observer",
          "passed": true,
          "source": "semantic_judge",
          "rationale": "The assistant achieved the verified outcome without violating safety or authorization constraints.",
          "failed_checks": [],
          "failed_procedure_steps": [
            "availability",
            "reservation",
            "verified_state",
            "confirm_result"
          ],
          "procedure_is_diagnostic": true
        }
      },
      "validity": {
        "scope": "caller_responses_delegation",
        "status": "valid",
        "target_metrics_eligible": true,
        "reason": null,
        "unexpected_delegation_count": 0,
        "delegation_ids": []
      }
    }
  ]
}
