{
  "task_id": "consumerfinance_014",
  "domain": "CFPB_COMPLAINT_DATABASE",
  "autonomy_type": "ordered table",
  "oracle_output_cardinality": 4,
  "instruction": "我正在起草一份关于加密货币投诉批次的核心因素逐层分析报告，需要你协助我通过 CFPB 投诉数据库，按步骤验证这波投诉高峰的真实成因。\n首先，我们需要界定分析的底层范围。请根据 CFPB 现行的产品与 Issue 口径文档，确认“Virtual currency”目前归属的产品大类（Product bucket）。在确定归属后，需将时间范围严格锁定在 2020-01-01 至 2022-08-26 之间，且仅针对该产品大类下的“Virtual currency”子类，提取出这个初始投诉群体的原始峰值月份（BaselinePeak）及其总投诉量。\n其次，我们需要通过层层剔除核心影响因素来推演峰值的转移情况。第一轮，需在上述原始峰值月里识别出出现频次最高的“Company public response”取值，并从整个初始群体中彻底移除该取值对应的所有投诉，从而获取重新计算后的第二个峰值月（AfterDominantPublicResponse）。第二轮，需在第二个峰值月里识别出出现频次最高的“Issue”类型，从上一轮残余数据中将其剔除，以锁定第三个涌现出的峰值月（AfterDominantPublicResponseAndIssue）。第三轮，需在第三个峰值月里识别出被投诉最多的单一公司（Company），并在剔除该公司后进行终轮统计，沉淀出最终的残余峰值月（AfterDominantPublicResponseIssueAndCompany）。\n再次，针对上述每一个识别出的特征月份，我们需要提取详尽的数据。对于这四个投诉簇（Cluster），除了必须获取该月总投诉量外，还需分别查实：本轮新增的移除维度及其具体的排除值，以及该排除值在它原本作为“参照来源”的那个月份里出现的次数。同时，还需明确该月份内投诉量最高的州及其投诉量、该州内被投诉最多的公司及其投诉规模，以及该州内投诉量最高的公司在该州、该特征月内收到的最早一条投诉的完整信息（包含投诉编号、日期、Issue 和 Sub-issue）。\n最后，在完成上述所有维度的排查与信息梳理后，若将这四个投诉簇的所有提取参数，严格按照 BaselinePeak、AfterDominantPublicResponse、AfterDominantPublicResponseAndIssue、AfterDominantPublicResponseIssueAndCompany 的先后顺序，分别拼接成四行没有任何空格的字符串（每行字段顺序严格固定为：<Cluster>|<Month>|<MonthComplaints>|<NewExclusionType>|<NewExclusionValue>|<NewExclusionCountInSourcePeak>|<TopState>|<StateComplaints>|<TopCompany>|<CompanyComplaints>|<EarliestComplaintID>|<EarliestDate>|<Issue>|<SubIssue>），最终生成的这组数据记录内容是什么？若出现并列，一律按字典序升序选择；日期相同的最早投诉，再按 Complaint ID 数值升序选择。所有空值先统一为 NONE 后再排序。",
  "start_url": "https://www.consumerfinance.gov/data-research/consumer-complaints/",
  "output_format": "严格按照 BaselinePeak、AfterDominantPublicResponse、AfterDominantPublicResponseAndIssue、AfterDominantPublicResponseIssueAndCompany 的先后顺序，分别拼接成四行没有任何空格的字符串；每行字段顺序严格固定为：<Cluster>|<Month>|<MonthComplaints>|<NewExclusionType>|<NewExclusionValue>|<NewExclusionCountInSourcePeak>|<TopState>|<StateComplaints>|<TopCompany>|<CompanyComplaints>|<EarliestComplaintID>|<EarliestDate>|<Issue>|<SubIssue>。若出现并列，一律按字典序升序选择；日期相同的最早投诉，再按 Complaint ID 数值升序选择。所有空值先统一为 NONE 后再排序。",
  "oracle_answer": "BaselinePeak|2021-05|336|NONE|NONE|0|CA|58|Coinbase, Inc.|21|4341451|2021-05-01|Money was not available when promised|NONE,AfterDominantPublicResponse|2022-07|31|Company public response|NONE|321|CA|6|BANK OF AMERICA, NATIONAL ASSOCIATION|4|5735227|2022-07-05|Other transaction problem|NONE,AfterDominantPublicResponseAndIssue|2021-10|13|Issue|Fraud or scam|26|TX|2|BANK OF AMERICA, NATIONAL ASSOCIATION|1|4821611|2021-10-18|Other transaction problem|NONE,AfterDominantPublicResponseIssueAndCompany|2021-06|8|Company|Foris DAX, Inc.|5|NC|3|TRUIST FINANCIAL CORPORATION|2|4462160|2021-06-15|Money was not available when promised|NONE",
  "metadata": {
    "State-Gated Retrieval": [
      "只看当前 product bucket 下 Virtual currency 子类、接收日期在 2020-01-01 至 2022-08-26 的 Complaint Database 记录。",
      "按题目要求依次确定 BaselinePeak、AfterDominantPublicResponse、AfterDominantIssue、AfterDominantCompany 四个阶段的峰值月。",
      "每一轮识别出的 dominant Company public response、dominant Issue、dominant Company，都必须回写到对应上游 residual cohort 中整体删除后，再重跑下一轮统计。"
    ],
    "dependency_type": "Data + Control",
    "intra_chain": true,
    "inter_chain": true,
    "data_dependency": [
      "current complaint-form taxonomy yields the product bucket for Virtual currency",
      "complaint records yield the baseline peak plus the dominant Company public response, dominant Issue, and dominant Company at successive residual-peak states",
      "final rows are produced only after each dominant value is applied to the full upstream residual cohort and then used to determine the next peak"
    ],
    "control_dependency": [
      "dominant values found in a peak month cannot be deleted only inside that month; each must be propagated to the full upstream cohort",
      "each new residual peak depends on the prior round's global rewrite, so local month-by-month deletion gives the wrong chain",
      "downstream counts and exemplar complaints are valid only after the full three-stage waterfall is executed"
    ],
    "freeze": {
      "historical_window": "Complaint Database records received from 2020-01-01 through 2022-08-26 under the current product-bucket mapping for Virtual currency"
    },
    "answer_type": "multi-row ordered table"
  },
  "rubric": {
    "inclusion_conditions": [
      "只看当前 product bucket 下 Virtual currency 子类、接收日期在 2020-01-01 至 2022-08-26 的 Complaint Database 记录。",
      "按题目要求依次确定 BaselinePeak、AfterDominantPublicResponse、AfterDominantIssue、AfterDominantCompany 四个阶段的峰值月。",
      "每一轮识别出的 dominant Company public response、dominant Issue、dominant Company，都必须回写到对应上游 residual cohort 中整体删除后，再重跑下一轮统计。"
    ],
    "exclusion_conditions": [
      "排除只在当前峰值月局部删记录、没有回写到整个上游 cohort 的结果。",
      "排除没有在每一轮 residual cohort 上重新寻找下一峰值月的结果。",
      "排除基于未整体重跑 downstream 统计得到的中间或最终字段。"
    ],
    "normalization": {
      "field_separator": "|",
      "record_separator": ",",
      "schema": [
        "Cluster",
        "Month",
        "MonthComplaints",
        "NewExclusionType",
        "NewExclusionValue",
        "NewExclusionCountInSourcePeak",
        "TopState",
        "StateComplaints",
        "TopCompany",
        "CompanyComplaints",
        "EarliestComplaintID",
        "EarliestDate",
        "Issue",
        "SubIssue"
      ],
      "cluster_order": [
        "BaselinePeak",
        "AfterDominantPublicResponse",
        "AfterDominantPublicResponseAndIssue",
        "AfterDominantPublicResponseIssueAndCompany"
      ],
      "date_format": {
        "Month": "YYYY-MM",
        "EarliestDate": "YYYY-MM-DD"
      },
      "NewExclusionType": "emit the canonical exclusion-type label from the source workflow; when no new exclusion is applied, emit `NONE`",
      "NewExclusionValue": "emit the canonical excluded value; when no value exists, emit `NONE`",
      "TopState": "use the uppercase two-letter state abbreviation",
      "TopCompany": "use the company display name as shown in the source data",
      "tie_breaks": "normalize empty values to `NONE` before sorting; for other ties use lexicographic order ascending, and for the earliest complaint use EarliestDate ascending then Complaint ID ascending",
      "SubIssue": "emit `NONE` when the complaint row has no sub-issue value"
    }
  }
}