{
  "note": "经人工整理的可见轨迹摘录；不是隐藏推理或连续录屏。部分GAgent持久化调用参数与结果原本已有截断。",
  "cases": [
    {
      "id": "device",
      "domain": "device",
      "title": "两个设置，是否真的改好了？",
      "short": "锁屏与自动亮度",
      "challenge": "状态核验 · 修改范围",
      "insight": "设置页打开、手势返回成功，都不能替代目标状态的独立读回。",
      "limitations": "配置值达标不等于实测两分钟后屏幕必然熄灭；设备原有“充电不熄屏”设置保持。豆包使用已有主会话，权限与历史上下文不完全匹配。",
      "ga": {
        "run": "ga_device_focus_base",
        "outcome": "partial",
        "label": "部分完成",
        "result": "自动亮度关闭；锁屏仍为30分钟。",
        "metric": "14轮 / 22次存储调用",
        "steps": [
          {
            "title": "确认已有操作能力",
            "detail": "基线出现多次接口与环境探测；旧设备记忆与当前设备不一致。",
            "evidence": "持久化上下文",
            "image": null,
            "observation": null
          },
          {
            "title": "进入显示设置",
            "detail": "观察到自动锁屏仍为30分钟。此截图记录运行中状态，不是最终状态。",
            "evidence": "采样截图",
            "image": "media/ga-device.png",
            "observation": "观察器 +120.435秒"
          },
          {
            "title": "预算结束后读回",
            "detail": "最终 screen_off_timeout=1800000，screen_brightness_mode=0；两项仅完成一项。",
            "evidence": "独立系统读数",
            "image": null,
            "observation": null
          }
        ],
        "artifact": null
      },
      "db": {
        "run": "db_device_focus_main",
        "outcome": "pass",
        "label": "目标状态达标",
        "result": "锁屏120000毫秒；自动亮度关闭。",
        "metric": "UI自述33秒；非统一计时",
        "steps": [
          {
            "title": "从手机任务主会话进入",
            "detail": "新对话存在模式资格限制；正式运行使用支持手机操作的主会话。",
            "evidence": "模式检查与提交记录",
            "image": null,
            "observation": null
          },
          {
            "title": "界面报告完成",
            "detail": "首次采到完成界面约在观察器第44.633秒。界面的33秒是产品自述。",
            "evidence": "采样截图",
            "image": "media/db-device.png",
            "observation": "观察器 +44.633秒"
          },
          {
            "title": "独立读取最终设置",
            "detail": "锁屏120000毫秒、自动亮度0；字体、铃声、媒体音量和勿扰字段保持。",
            "evidence": "独立系统读数",
            "image": null,
            "observation": null
          }
        ],
        "artifact": null
      },
      "query": "接下来我要对着纸质资料用手机做检查。请把自动锁屏时间改成2分钟，并关闭自动亮度；字体大小和声音设置保持原样。完成后告诉我这两项的实际状态。",
      "query_sha256": "4fa314e5068e74248d73480d0aa8a5b32b0abf3ad2d3656d373efde9707cfde8"
    },
    {
      "id": "travel",
      "domain": "travel",
      "title": "一趟赶得上的出差，四个候选。",
      "short": "差旅候选筛选",
      "challenge": "时间约束 · 预算 · 可用性",
      "insight": "最终方案正确，并不意味着过程中的每一句说明都正确。最终文件需要单独核对。",
      "limitations": "固定报价题诊断条件核算与附件交付，不代表实时查票或跨App导航。豆包继承主会话；两产品计时来源不同。",
      "ga": {
        "run": "ga_travel_deadline",
        "outcome": "pass",
        "label": "方案与附件正确",
        "result": "A车 · 11:10出门 · 13:30到场 · ¥231",
        "metric": "6轮 / 11次存储调用",
        "steps": [
          {
            "title": "读取规程并核算约束",
            "detail": "分别检查到场上限、出门下限、二等座余票和总预算。",
            "evidence": "持久化上下文",
            "image": null,
            "observation": null
          },
          {
            "title": "淘汰三个不可行候选",
            "detail": "B到场13:55过晚；C需10:45离家过早；D二等座无票。",
            "evidence": "持久化上下文与Markdown",
            "image": null,
            "observation": null
          },
          {
            "title": "生成并交付比较文件",
            "detail": "实际Markdown附件3632字节，页面文件卡与原字节均已核对。",
            "evidence": "交付截图与文件",
            "image": "media/ga-travel.png",
            "observation": null
          }
        ],
        "artifact": {
          "url": "downloads/ga_travel_deadline.md",
          "bytes": 3632,
          "sha256": "91326e195b860ad4a12b3dd3d85856155cd9bb6891a084ad371fab184ec2540d",
          "checks": {
            "A_recommended": true,
            "A_latest_departure_11_10": true,
            "A_arrival_13_30": true,
            "A_total_231": true,
            "B_arrival_13_55_rejected": true,
            "C_departure_10_45_rejected": true,
            "D_no_second_class_seat_rejected": true,
            "source_snapshot_only_in_explanation": true
          },
          "format": "MD"
        }
      },
      "db": {
        "run": "db_travel_deadline",
        "outcome": "pass",
        "label": "方案与附件正确",
        "result": "A车 · 11:10出门 · 13:30到场 · ¥231",
        "metric": "附件首见约提交后80秒",
        "steps": [
          {
            "title": "分析时间和预算条件",
            "detail": "公开面板展示候选筛选过程；部分说明的离家时间措辞有误。",
            "evidence": "公开过程面板",
            "image": "media/db-travel-plan.png",
            "observation": null
          },
          {
            "title": "复核时间换算",
            "detail": "计算异常后重新核对时间，最终文件正确排除B、C、D。",
            "evidence": "公开过程面板",
            "image": "media/db-travel-recheck.png",
            "observation": null
          },
          {
            "title": "交付可打开的Markdown",
            "detail": "真实下载文件2196字节，A方案的时间、金额和淘汰理由正确。",
            "evidence": "实际附件界面",
            "image": "media/db-travel.png",
            "observation": null
          }
        ],
        "artifact": {
          "url": "downloads/db_travel_deadline.md",
          "bytes": 2196,
          "sha256": "12de743d1506ef14966aa263293e72adfc0707a733bedf8b88ae27f69cc0774e",
          "checks": {
            "A_recommended": true,
            "A_latest_departure_11_10": true,
            "A_arrival_13_30": true,
            "A_total_231": true,
            "B_arrival_13_55_rejected": true,
            "C_departure_10_45_rejected": true,
            "D_no_second_class_seat_rejected": true,
            "source_snapshot_only_in_explanation": true
          },
          "format": "MD"
        }
      },
      "query": "请帮我定明天出差的候选方案，只用下面这份已核实的报价快照，不用联网或预订。我明天14:00在杭州见客户，要提前20分钟到会场，11:00前不能离开家。家到上海出发站45分钟，进站预留25分钟，打车35元。候选：A车12:20出发、13:10到杭州，票178元，二等座有票，到会场20分钟/18元；B车12:35出发、13:25到，票152元，有票，到会场30分钟/22元；C车11:55出发、12:45到，票166元，有票，到会场20分钟/18元；D车12:15出发、13:05到，票148元，但二等座无票，到会场20分钟/18元。去程总预算250元。请推荐满足所有条件的方案，写清最晚出门时间、到会场时间和去程总额，并逐个解释其他候选不合适的原因。做成一个可打开的Markdown附件，文件名R30出差比较.md；不要创建日历或闹钟。",
      "query_sha256": "0bd53ccab9ec10d40017c8647f3c1b33bb4d59799bf7dbe6bbfeaf6ce00fd963"
    },
    {
      "id": "finance",
      "domain": "finance",
      "title": "同样的金额，不一定是同一笔钱。",
      "short": "去重与退款CSV",
      "challenge": "身份判定 · 退款状态 · CSV结构",
      "insight": "答案总额之外，检查订单身份、未完成退款和带逗号字段，才能知道产物能否继续使用。",
      "limitations": "使用合成流水，不访问真实支付账户。豆包约227秒是研究者首见结果的上界，视口未跟随，不能当作生成耗时。",
      "ga": {
        "run": "ga_finance_refund",
        "outcome": "pass",
        "label": "20项产物检查通过",
        "result": "支付293.30 − 已退64.90 = 净额228.40",
        "metric": "2轮 / 1次存储调用",
        "steps": [
          {
            "title": "按订单身份去重",
            "detail": "T01重复仅计一次；T02虽然同额仍独立保留。只抵扣成功退款。",
            "evidence": "持久化上下文",
            "image": null,
            "observation": null
          },
          {
            "title": "写文件并读回",
            "detail": "使用CSV写入器，保留UTF-8 BOM。含逗号商户正确引用，全额退款行保留。",
            "evidence": "存储工具调用与原文件",
            "image": null,
            "observation": null
          },
          {
            "title": "打开真实附件预览",
            "detail": "184字节原件通过20项独立检查，手机预览为五列且商户没有错列。",
            "evidence": "附件预览",
            "image": "media/ga-finance.png",
            "observation": null
          }
        ],
        "artifact": {
          "url": "downloads/ga_finance_refund.csv",
          "bytes": 184,
          "sha256": "0e374c13dd2295d6d59359c1a28104a36dd899cad666e97af14e88570f352973",
          "checks": {
            "header": true,
            "row_count": true,
            "rectangular": true,
            "order_ids": true,
            "total_last": true,
            "merchants": true,
            "merchant_with_comma": true,
            "total_merchant_empty": true,
            "T01.paid": true,
            "T01.refunded": true,
            "T01.net": true,
            "T02.paid": true,
            "T02.refunded": true,
            "T02.net": true,
            "T03.paid": true,
            "T03.refunded": true,
            "T03.net": true,
            "TOTAL.paid": true,
            "TOTAL.refunded": true,
            "TOTAL.net": true
          },
          "format": "CSV"
        }
      },
      "db": {
        "run": "db_finance_refund",
        "outcome": "pass",
        "label": "20项产物检查通过",
        "result": "支付293.30 − 已退64.90 = 净额228.40",
        "metric": "首见时间上界，非生成耗时",
        "steps": [
          {
            "title": "生成UTF-8 CSV",
            "detail": "公开过程表明已生成带BOM的CSV，并进入结构检查。",
            "evidence": "公开过程面板",
            "image": "media/db-finance-build.png",
            "observation": null
          },
          {
            "title": "修正校验参数",
            "detail": "过程显示先修正JSON参数格式，再改用单元格定位检查合计与关键数值。",
            "evidence": "公开过程面板",
            "image": "media/db-finance-check.png",
            "observation": null
          },
          {
            "title": "交付并解释排除依据",
            "detail": "179字节原件通过20项独立检查；净额、退款口径和带逗号商户均正确。",
            "evidence": "最终回复与实际附件",
            "image": "media/db-finance.png",
            "observation": null
          }
        ],
        "artifact": {
          "url": "downloads/db_finance_refund.csv",
          "bytes": 179,
          "sha256": "32e24751fc06b85affbf6327dbfc1b1f0aac2347bdb59b5c44a82b54be6bebb4",
          "checks": {
            "header": true,
            "row_count": true,
            "rectangular": true,
            "order_ids": true,
            "total_last": true,
            "merchants": true,
            "merchant_with_comma": true,
            "total_merchant_empty": true,
            "T01.paid": true,
            "T01.refunded": true,
            "T01.net": true,
            "T02.paid": true,
            "T02.refunded": true,
            "T02.net": true,
            "T03.paid": true,
            "T03.refunded": true,
            "T03.net": true,
            "TOTAL.paid": true,
            "TOTAL.refunded": true,
            "TOTAL.net": true
          },
          "format": "CSV"
        }
      },
      "query": "我在整理9月出差报销，请只根据下面的流水做一个可打开的UTF-8 CSV附件，文件名R30报销31.csv。\n金额都是人民币；按订单编号去重，同编号完全相同的支付记录只记一次，不同订单即使金额相同也分别计入。只扣除已成功退款，退款申请中不扣。全额退款订单保留一行，净支出为0。\n流水：\n订单T01，商户“车站,便利店”，支付成功128.40元；同一笔T01支付成功128.40元在导出表又出现一次；T01退款成功28.40元。\n订单T02，商户“城际车票”，支付成功128.40元；T02退款申请中10.00元。\n订单T03，商户“会议午餐”，支付成功36.50元；T03退款成功36.50元。\n订单T04，商户“会场接驳”，支付失败18.00元。\nCSV只保留T01、T02、T03三条订单及最后一条TOTAL，表头严格为order_id,merchant,paid,refunded,net。金额保留两位小数，TOTAL行merchant留空，后三列分别为合计。请同时说明排除重复和未成功交易的依据，不修改任何账单原件，也不用访问支付App。",
      "query_sha256": "5b351d9087628f00df23c216c1dc48d96e065ac4112477199c85d7fb4c47c543"
    },
    {
      "id": "schedule",
      "domain": "productivity",
      "title": "50分钟会议，还要留出两侧缓冲。",
      "short": "会议缓冲与时区ICS",
      "challenge": "区间求解 · 时区语义 · 附件交付",
      "insight": "文件可以解析，不代表时间语义正确。跨时区使用时，浮动时间不对应同一绝对时刻。",
      "limitations": "豆包观察器最后采样在提交后293.371秒，严格第300秒状态未知；后续取得的附件可独立判断时区与命名。没有导入日历或发送邀请。",
      "ga": {
        "run": "ga_schedule_buffers",
        "outcome": "pass",
        "label": "10项产物检查通过",
        "result": "13:50–14:40 · Asia/Shanghai · 上午无解",
        "metric": "3轮 / 4次存储调用",
        "steps": [
          {
            "title": "核验工作区间和既有会议",
            "detail": "先读取规程；对所有已有日程施加双侧10分钟间隔。",
            "evidence": "持久化上下文",
            "image": null,
            "observation": null
          },
          {
            "title": "计算可行档并生成ICS",
            "detail": "上午最大有效空档不足50分钟；下午首个可行档为13:50至14:40。",
            "evidence": "持久化上下文与原文件",
            "image": null,
            "observation": null
          },
          {
            "title": "读回带时区的原件",
            "detail": "614字节ICS包含TZID与VTIMEZONE，无提醒或参会人，10项检查通过。",
            "evidence": "真实ICS预览",
            "image": "media/ga-schedule-preview.png",
            "observation": null
          }
        ],
        "artifact": {
          "url": "downloads/ga_schedule_buffers.ics",
          "bytes": 614,
          "sha256": "8c7192722c7ff49d541703d4ca04c0cc7859b1d7027c99a3281bc1272605702d",
          "checks": {
            "one_event": true,
            "no_alarms": true,
            "calendar_version": true,
            "preferred_start": true,
            "preferred_end": true,
            "summary": true,
            "description": true,
            "no_attendees": true,
            "no_recurrence": true,
            "identity_and_timestamp": true
          },
          "format": "ICS"
        }
      },
      "db": {
        "run": "db_schedule_buffers",
        "outcome": "partial",
        "label": "排期正确，附件有偏差",
        "result": "时间正确；ICS为浮动时间，文件名偏离要求",
        "metric": "第300秒状态未知",
        "steps": [
          {
            "title": "得到正确的会议区间",
            "detail": "确认13:40加10分钟后开始，持续50分钟；上午无可用备选。",
            "evidence": "公开过程面板",
            "image": "media/db-schedule-check.png",
            "observation": null
          },
          {
            "title": "额外制作时间轴",
            "detail": "读取可视化规程，寻找Chromium并渲染、修复时间轴。题目并未要求该图。",
            "evidence": "公开过程面板",
            "image": "media/db-schedule-extra.png",
            "observation": null
          },
          {
            "title": "后来取得附件并独立检查",
            "detail": "308字节ICS可解析，但起止时间没有TZID或UTC后缀；文件名额外带日期。",
            "evidence": "实际交付界面与ICS",
            "image": "media/db-schedule.png",
            "observation": null
          }
        ],
        "artifact": {
          "url": "downloads/db_schedule_buffers.ics",
          "bytes": 308,
          "sha256": "49e37262fb196a4355c1010e770069b44ee60ac19c9e037f4baebd383060574b",
          "checks": {
            "one_event": true,
            "no_alarms": true,
            "calendar_version": true,
            "preferred_start": false,
            "preferred_end": false,
            "summary": true,
            "description": true,
            "no_attendees": true,
            "no_recurrence": true,
            "identity_and_timestamp": true
          },
          "format": "ICS"
        }
      },
      "query": "帮我找出2026年10月8日的会议时间并生成可导入日历的R30方案评审.ics附件，暂时不要向任何人发送邀请，也不要直接写入我的日历。下面所有时间都是上海时间UTC+8，不用联网查节假日。\n我和同事都能开会的范围是09:00—12:00和13:00—17:00，新会议“方案评审”需要连续50分钟，优先选下午最早的可行时间。上午我已有09:00—10:00访谈、11:00—11:30周会；下午我已有13:00—13:40客户电话、15:00—15:30站会。同事下午已有14:50—15:20培训。新会议与每一场既有日程之间都需要至少10分钟间隔，不能移动已有日程。上午若有可行时间也请列出最早一档作备选。\n请解释你选的起止时间和备选时间，附件只含首选的一个事件，不设提醒、不加参会人，会议说明写“会前查看方案第2版”。",
      "query_sha256": "4fa841bb1d620fac85e3e303a2ab548666d4cba710b2524c81600108a5cd01d7"
    }
  ]
}