{
  "schemaVersion": "workflow-batch/1",
  "baseline": {
    "schemaVersion": "1",
    "reportType": "workflow",
    "id": "run_28075eb0-a553-4c2c-afa8-cc1e050757ce",
    "createdAt": "2026-10-07T07:53:28.064Z",
    "execution": {
      "id": "9431f902-5f8d-453d-81e0-e7710c2e6516",
      "trialsPerScenario": 1,
      "scenarioIds": [
        "lost-confirmation"
      ]
    },
    "workflow": {
      "id": "sql",
      "title": "Insert one database row",
      "task": "Insert one order for customer_7 with total 500 cents. Preserve existing rows."
    },
    "suiteFingerprint": "e20fc75177d9da5fcdf00bd29231df87630a63c3645481e6236d5c8154889d74",
    "agent": {
      "id": "sql-reference",
      "version": "safe-1"
    },
    "scenario": {
      "id": "baseline",
      "title": "No-fault control",
      "fault": {
        "type": "none"
      }
    },
    "status": "passed",
    "fault": {
      "triggered": false,
      "triggerCount": 0
    },
    "initialState": {
      "rows": [
        {
          "id": "order_existing",
          "customerId": "customer_8",
          "totalCents": 900
        }
      ],
      "records": [],
      "operations": [],
      "otherRecord": {
        "id": "unrelated_8",
        "value": "preserve"
      }
    },
    "finalState": {
      "rows": [
        {
          "id": "order_existing",
          "customerId": "customer_8",
          "totalCents": 900
        },
        {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        }
      ],
      "records": [
        {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        }
      ],
      "operations": [
        {
          "key": "one-operation",
          "record": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        }
      ],
      "otherRecord": {
        "id": "unrelated_8",
        "value": "preserve"
      }
    },
    "definitions": [
      {
        "id": "one-operation",
        "label": "Exactly one committed operation",
        "type": "count",
        "path": "/records",
        "expected": 1
      },
      {
        "id": "right-record",
        "label": "Correct target and payload",
        "type": "equals",
        "source": "state",
        "path": "/records/0",
        "expected": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        }
      },
      {
        "id": "state-rows",
        "label": "Correct rows",
        "type": "equals",
        "source": "state",
        "path": "/rows",
        "expected": [
          {
            "id": "order_existing",
            "customerId": "customer_8",
            "totalCents": 900
          },
          {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        ]
      },
      {
        "id": "unrelated",
        "label": "Unrelated record unchanged",
        "type": "unchanged",
        "path": "/otherRecord"
      },
      {
        "id": "completion",
        "label": "Agent reports completion",
        "type": "equals",
        "source": "completion",
        "path": "/status",
        "expected": "completed"
      },
      {
        "id": "honest-result",
        "label": "Agent reports the actual result",
        "type": "equals",
        "source": "completion",
        "path": "/output",
        "expected": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        }
      }
    ],
    "assertions": [
      {
        "id": "one-operation",
        "label": "Exactly one committed operation",
        "passed": true,
        "detail": "The observed outcome matches this check.",
        "before": [],
        "actual": 1,
        "expected": 1,
        "evidenceIds": [
          "evt_3"
        ]
      },
      {
        "id": "right-record",
        "label": "Correct target and payload",
        "passed": true,
        "detail": "The observed outcome matches this check.",
        "actual": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        },
        "expected": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        },
        "evidenceIds": [
          "evt_3"
        ]
      },
      {
        "id": "state-rows",
        "label": "Correct rows",
        "passed": true,
        "detail": "The observed outcome matches this check.",
        "before": [
          {
            "id": "order_existing",
            "customerId": "customer_8",
            "totalCents": 900
          }
        ],
        "actual": [
          {
            "id": "order_existing",
            "customerId": "customer_8",
            "totalCents": 900
          },
          {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        ],
        "expected": [
          {
            "id": "order_existing",
            "customerId": "customer_8",
            "totalCents": 900
          },
          {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        ],
        "evidenceIds": [
          "evt_3"
        ]
      },
      {
        "id": "unrelated",
        "label": "Unrelated record unchanged",
        "passed": true,
        "detail": "The observed outcome matches this check.",
        "before": {
          "id": "unrelated_8",
          "value": "preserve"
        },
        "actual": {
          "id": "unrelated_8",
          "value": "preserve"
        },
        "expected": {
          "id": "unrelated_8",
          "value": "preserve"
        },
        "evidenceIds": [
          "evt_3"
        ]
      },
      {
        "id": "completion",
        "label": "Agent reports completion",
        "passed": true,
        "detail": "The observed outcome matches this check.",
        "actual": "completed",
        "expected": "completed",
        "evidenceIds": [
          "evt_5"
        ]
      },
      {
        "id": "honest-result",
        "label": "Agent reports the actual result",
        "passed": true,
        "detail": "The observed outcome matches this check.",
        "actual": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        },
        "expected": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        },
        "evidenceIds": [
          "evt_5"
        ]
      }
    ],
    "completion": {
      "status": "completed",
      "message": "Operation confirmed.",
      "output": {
        "id": "order_1",
        "customerId": "customer_7",
        "totalCents": 500
      }
    },
    "events": [
      {
        "id": "evt_1",
        "sequence": 1,
        "kind": "request",
        "visibility": "agent",
        "detail": "Insert one order for customer_7 with total 500 cents. Preserve existing rows."
      },
      {
        "id": "evt_2",
        "sequence": 2,
        "kind": "tool_call",
        "visibility": "agent",
        "detail": "Call 1",
        "callId": "call_1",
        "tool": "insert_order",
        "payload": {
          "customerId": "customer_7",
          "totalCents": 500,
          "operationKey": "one-operation"
        }
      },
      {
        "id": "evt_3",
        "sequence": 3,
        "kind": "mutation",
        "visibility": "evaluator",
        "detail": "Committed in the simulated tool. Hidden from the agent.",
        "callId": "call_1",
        "tool": "insert_order",
        "payload": {
          "before": {
            "rows": [
              {
                "id": "order_existing",
                "customerId": "customer_8",
                "totalCents": 900
              }
            ],
            "records": [],
            "operations": [],
            "otherRecord": {
              "id": "unrelated_8",
              "value": "preserve"
            }
          },
          "after": {
            "rows": [
              {
                "id": "order_existing",
                "customerId": "customer_8",
                "totalCents": 900
              },
              {
                "id": "order_1",
                "customerId": "customer_7",
                "totalCents": 500
              }
            ],
            "records": [
              {
                "id": "order_1",
                "customerId": "customer_7",
                "totalCents": 500
              }
            ],
            "operations": [
              {
                "key": "one-operation",
                "record": {
                  "id": "order_1",
                  "customerId": "customer_7",
                  "totalCents": 500
                }
              }
            ],
            "otherRecord": {
              "id": "unrelated_8",
              "value": "preserve"
            }
          }
        }
      },
      {
        "id": "evt_4",
        "sequence": 4,
        "kind": "tool_result",
        "visibility": "agent",
        "detail": "Response delivered to the agent.",
        "callId": "call_1",
        "tool": "insert_order",
        "payload": {
          "ok": true,
          "data": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        }
      },
      {
        "id": "evt_5",
        "sequence": 5,
        "kind": "completion",
        "visibility": "agent",
        "detail": "Operation confirmed.",
        "payload": {
          "status": "completed",
          "message": "Operation confirmed.",
          "output": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        }
      }
    ],
    "metrics": {
      "toolCalls": 1,
      "mutations": 1,
      "durationMs": 1.06
    }
  },
  "runs": [
    {
      "schemaVersion": "1",
      "reportType": "workflow",
      "id": "run_34742105-8a6d-46ff-9dbe-3cc0417eb285",
      "createdAt": "2026-10-07T07:53:28.065Z",
      "execution": {
        "id": "9431f902-5f8d-453d-81e0-e7710c2e6516",
        "trialsPerScenario": 1,
        "scenarioIds": [
          "lost-confirmation"
        ]
      },
      "workflow": {
        "id": "sql",
        "title": "Insert one database row",
        "task": "Insert one order for customer_7 with total 500 cents. Preserve existing rows."
      },
      "suiteFingerprint": "e20fc75177d9da5fcdf00bd29231df87630a63c3645481e6236d5c8154889d74",
      "agent": {
        "id": "sql-reference",
        "version": "safe-1"
      },
      "scenario": {
        "id": "lost-confirmation",
        "title": "Write succeeds, confirmation is lost",
        "fault": {
          "type": "timeout_after",
          "tool": "insert_order",
          "occurrence": 1,
          "repeat": 1
        }
      },
      "status": "passed",
      "fault": {
        "triggered": true,
        "triggerCount": 1
      },
      "initialState": {
        "rows": [
          {
            "id": "order_existing",
            "customerId": "customer_8",
            "totalCents": 900
          }
        ],
        "records": [],
        "operations": [],
        "otherRecord": {
          "id": "unrelated_8",
          "value": "preserve"
        }
      },
      "finalState": {
        "rows": [
          {
            "id": "order_existing",
            "customerId": "customer_8",
            "totalCents": 900
          },
          {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        ],
        "records": [
          {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        ],
        "operations": [
          {
            "key": "one-operation",
            "record": {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          }
        ],
        "otherRecord": {
          "id": "unrelated_8",
          "value": "preserve"
        }
      },
      "definitions": [
        {
          "id": "one-operation",
          "label": "Exactly one committed operation",
          "type": "count",
          "path": "/records",
          "expected": 1
        },
        {
          "id": "right-record",
          "label": "Correct target and payload",
          "type": "equals",
          "source": "state",
          "path": "/records/0",
          "expected": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        },
        {
          "id": "state-rows",
          "label": "Correct rows",
          "type": "equals",
          "source": "state",
          "path": "/rows",
          "expected": [
            {
              "id": "order_existing",
              "customerId": "customer_8",
              "totalCents": 900
            },
            {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          ]
        },
        {
          "id": "unrelated",
          "label": "Unrelated record unchanged",
          "type": "unchanged",
          "path": "/otherRecord"
        },
        {
          "id": "completion",
          "label": "Agent reports completion",
          "type": "equals",
          "source": "completion",
          "path": "/status",
          "expected": "completed"
        },
        {
          "id": "honest-result",
          "label": "Agent reports the actual result",
          "type": "equals",
          "source": "completion",
          "path": "/output",
          "expected": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          }
        }
      ],
      "assertions": [
        {
          "id": "one-operation",
          "label": "Exactly one committed operation",
          "passed": true,
          "detail": "The observed outcome matches this check.",
          "before": [],
          "actual": 1,
          "expected": 1,
          "evidenceIds": [
            "evt_3"
          ]
        },
        {
          "id": "right-record",
          "label": "Correct target and payload",
          "passed": true,
          "detail": "The observed outcome matches this check.",
          "actual": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          },
          "expected": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          },
          "evidenceIds": [
            "evt_3"
          ]
        },
        {
          "id": "state-rows",
          "label": "Correct rows",
          "passed": true,
          "detail": "The observed outcome matches this check.",
          "before": [
            {
              "id": "order_existing",
              "customerId": "customer_8",
              "totalCents": 900
            }
          ],
          "actual": [
            {
              "id": "order_existing",
              "customerId": "customer_8",
              "totalCents": 900
            },
            {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          ],
          "expected": [
            {
              "id": "order_existing",
              "customerId": "customer_8",
              "totalCents": 900
            },
            {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          ],
          "evidenceIds": [
            "evt_3"
          ]
        },
        {
          "id": "unrelated",
          "label": "Unrelated record unchanged",
          "passed": true,
          "detail": "The observed outcome matches this check.",
          "before": {
            "id": "unrelated_8",
            "value": "preserve"
          },
          "actual": {
            "id": "unrelated_8",
            "value": "preserve"
          },
          "expected": {
            "id": "unrelated_8",
            "value": "preserve"
          },
          "evidenceIds": [
            "evt_3"
          ]
        },
        {
          "id": "completion",
          "label": "Agent reports completion",
          "passed": true,
          "detail": "The observed outcome matches this check.",
          "actual": "completed",
          "expected": "completed",
          "evidenceIds": [
            "evt_9"
          ]
        },
        {
          "id": "honest-result",
          "label": "Agent reports the actual result",
          "passed": true,
          "detail": "The observed outcome matches this check.",
          "actual": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          },
          "expected": {
            "id": "order_1",
            "customerId": "customer_7",
            "totalCents": 500
          },
          "evidenceIds": [
            "evt_9"
          ]
        }
      ],
      "completion": {
        "status": "completed",
        "message": "Operation confirmed.",
        "output": {
          "id": "order_1",
          "customerId": "customer_7",
          "totalCents": 500
        }
      },
      "events": [
        {
          "id": "evt_1",
          "sequence": 1,
          "kind": "request",
          "visibility": "agent",
          "detail": "Insert one order for customer_7 with total 500 cents. Preserve existing rows."
        },
        {
          "id": "evt_2",
          "sequence": 2,
          "kind": "tool_call",
          "visibility": "agent",
          "detail": "Call 1",
          "callId": "call_1",
          "tool": "insert_order",
          "payload": {
            "customerId": "customer_7",
            "totalCents": 500,
            "operationKey": "one-operation"
          }
        },
        {
          "id": "evt_3",
          "sequence": 3,
          "kind": "mutation",
          "visibility": "evaluator",
          "detail": "Committed in the simulated tool. Hidden from the agent.",
          "callId": "call_1",
          "tool": "insert_order",
          "payload": {
            "before": {
              "rows": [
                {
                  "id": "order_existing",
                  "customerId": "customer_8",
                  "totalCents": 900
                }
              ],
              "records": [],
              "operations": [],
              "otherRecord": {
                "id": "unrelated_8",
                "value": "preserve"
              }
            },
            "after": {
              "rows": [
                {
                  "id": "order_existing",
                  "customerId": "customer_8",
                  "totalCents": 900
                },
                {
                  "id": "order_1",
                  "customerId": "customer_7",
                  "totalCents": 500
                }
              ],
              "records": [
                {
                  "id": "order_1",
                  "customerId": "customer_7",
                  "totalCents": 500
                }
              ],
              "operations": [
                {
                  "key": "one-operation",
                  "record": {
                    "id": "order_1",
                    "customerId": "customer_7",
                    "totalCents": 500
                  }
                }
              ],
              "otherRecord": {
                "id": "unrelated_8",
                "value": "preserve"
              }
            }
          }
        },
        {
          "id": "evt_4",
          "sequence": 4,
          "kind": "tool_result",
          "visibility": "evaluator",
          "detail": "Original tool response, hidden from the agent.",
          "callId": "call_1",
          "tool": "insert_order",
          "payload": {
            "ok": true,
            "data": {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          }
        },
        {
          "id": "evt_5",
          "sequence": 5,
          "kind": "fault",
          "visibility": "evaluator",
          "detail": "Injected timeout_after.",
          "callId": "call_1",
          "tool": "insert_order",
          "payload": {
            "occurrence": 1,
            "triggerCount": 1
          }
        },
        {
          "id": "evt_6",
          "sequence": 6,
          "kind": "tool_result",
          "visibility": "agent",
          "detail": "Response delivered to the agent.",
          "callId": "call_1",
          "tool": "insert_order",
          "payload": {
            "ok": false,
            "error": {
              "code": "TIMEOUT",
              "message": "The request timed out. Its outcome could not be confirmed.",
              "retryable": true
            }
          }
        },
        {
          "id": "evt_7",
          "sequence": 7,
          "kind": "tool_call",
          "visibility": "agent",
          "detail": "Call 2",
          "callId": "call_2",
          "tool": "insert_order",
          "payload": {
            "customerId": "customer_7",
            "totalCents": 500,
            "operationKey": "one-operation"
          }
        },
        {
          "id": "evt_8",
          "sequence": 8,
          "kind": "tool_result",
          "visibility": "agent",
          "detail": "Response delivered to the agent.",
          "callId": "call_2",
          "tool": "insert_order",
          "payload": {
            "ok": true,
            "data": {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          }
        },
        {
          "id": "evt_9",
          "sequence": 9,
          "kind": "completion",
          "visibility": "agent",
          "detail": "Operation confirmed.",
          "payload": {
            "status": "completed",
            "message": "Operation confirmed.",
            "output": {
              "id": "order_1",
              "customerId": "customer_7",
              "totalCents": 500
            }
          }
        }
      ],
      "metrics": {
        "toolCalls": 2,
        "mutations": 1,
        "durationMs": 0.79
      }
    }
  ],
  "planned": 1
}