宜博 / lemonboss

{"id":"train-01","source":"real","provenance":"data/companies/e2e-demo/runs/run_9168bd24f46c/nodes/rn_6bcda7e1fd6a/attempt-0.json + contract.json 历史版本（remediation_note 记录的修改前状态）","excerpt":"六模块四接口一次交付契约，M-6 checks/check.mjs 作为验收依据，'三种宽度下无横向滚动'类结论句判据，'字号档位杂乱'无量化上限，无 content_fidelity_floor 兜底占位退化情形","scores":{"D1":4,"D2":1,"D3":4,"D4":5,"D5":5},"weighted_sum":68,"final_score":45,"cap_triggered":"D2=1 → 上限45","level":"不合格","label_note":"真实反面案例：被 planner-validator 实际打回重写。D2 命中全部4类缺陷模式（自证式验收/结论句代替实测值/风格描述无上限/占位可通过），触发封顶，与历史事实（确实被打回）一致。"}
{"id":"train-02","source":"real","provenance":"data/companies/e2e-demo/runs/run_c7d233619784/nodes/rn_4618bc2aa72d/contract.json","excerpt":"三模块四接口，scrollToTarget(targetId:string)->void 与 errors 同对象内嵌，data_shapes 给出 {id,title(非空),desc(非空)}，error_mapping 五条覆盖视口边界/换行/连点/JS禁用/极端窄屏","scores":{"D1":5,"D2":5,"D3":5,"D4":5,"D5":5},"weighted_sum":100,"final_score":100,"cap_triggered":null,"level":"优秀","label_note":"真实正面样板：一轮通过，无 remediation。签名+错误同对象、字段清单带约束、异常落点具体、无越界内容。"}
{"id":"train-03","source":"constructed","provenance":"构造样本，非真实案例，模拟 D1 违规","excerpt":"interfaces: [{id:'I-1', module:'M-1', signature:'exportOrders(params) -> 处理结果对象', errors:[]}]，全文再无对该'处理结果对象'的键/类型说明，异常处理单独写在正文'异常章节'笼统一句'导出失败会给出提示'","scores":{"D1":2,"D2":3,"D3":3,"D4":2,"D5":5},"weighted_sum":52,"final_score":52,"cap_triggered":null,"level":"待改进","label_note":"命中 agent.md 表#2'返回结果对象空话'与缺陷模式'错误行为拆到附录'，D1 命中2分档；连带 D4 因'导出失败会给出提示'无具体落点也扣分。"}
{"id":"train-04","source":"constructed","provenance":"构造样本，非真实案例，模拟 D2 单一违规：自证式验收","excerpt":"interfaces/data_shapes 均写实（键/类型/约束齐全），但契约末尾写'本模块交付时运行 npm run self-check，退出码0即视为验收通过'，验收判据完全依赖实现方自带脚本，无第三方可独立核对的判定语义","scores":{"D1":5,"D2":2,"D3":5,"D4":4,"D5":5},"weighted_sum":80,"final_score":50,"cap_triggered":"D2=2 → 上限45（此处加权和80>45，取45）","level":"不合格","label_note":"专门校准'自证式验收单独出现'时封顶规则是否会误伤其余维度全优的契约——结果：即便其余四维接近满分，D2命中核心判据也会被封顶到不合格，验证封顶规则确实按D2独立生效而非被稀释。"}
{"id":"train-05","source":"constructed","provenance":"构造样本，非真实案例，模拟 D2 单一违规：结论句代替实测值","excerpt":"验收相关判据写'页面在主流分辨率下显示正常，无溢出问题'，未给出具体分辨率清单与可复算的 scrollWidth/clientWidth 数值对或计算公式，其余接口签名与数据形态均具体完整","scores":{"D1":5,"D2":2,"D3":5,"D4":4,"D5":5},"weighted_sum":80,"final_score":45,"cap_triggered":"D2=2 → 上限45","level":"不合格","label_note":"结论句'显示正常'无法被第三方复算，命中D2缺陷模式'结论句代替实测值'，D2判2分（该处属核心判据，非次要边界样本）。"}
{"id":"train-06","source":"constructed","provenance":"构造样本，非真实案例，模拟 D3 违规：只写存起来","excerpt":"data_shapes: [{name:'用户导出记录', where:'数据库', shape:'保存导出历史'}]，无字段清单、无类型、无约束、无具体表/文件路径","scores":{"D1":5,"D2":4,"D3":1,"D4":4,"D5":5},"weighted_sum":78,"final_score":78,"cap_triggered":null,"level":"合格","label_note":"命中 agent.md 表#3 判不合格信号原句'只写\"存起来\"，没写存哪、什么形状、谁读'，D3判1分档（完全没有字段清单）。D2非核心问题所以不触发封顶，验证D3单独违规不会像D2那样拖累到不合格档，符合D2权重独立设计的预期。"}
{"id":"train-07","source":"constructed","provenance":"构造样本，非真实案例，模拟 D4 违规：漏落点","excerpt":"PRD 有 error_cases: [E-1空输入, E-2超长, E-3并发提交]，contract.json 的 error_mapping 只覆盖了 E-1/E-2，E-3（并发提交）在全文任何地方都找不到对应的 handled_by/how","scores":{"D1":5,"D2":4,"D3":5,"D4":3,"D5":5},"weighted_sum":84,"final_score":84,"cap_triggered":null,"level":"优秀","label_note":"命中 agent.md 表#4'PRD写了异常，SDD里找不到对应落点'，1条error_case缺落点对应D4三分档。用于校准：D4单项违规不足以让总分跌出优秀档，因为其权重(15)低于D2(30)，符合'D2是唯一必须封顶的核心维度'的设计意图。"}
{"id":"train-08","source":"constructed","provenance":"构造样本，非真实案例，模拟 D5 违规：越权吸收验收标准","excerpt":"contract.json 里出现'验收权重表：接口契约完整性占40%，视觉还原度占30%，性能占30%，及格线70分'——这类验收权重/及格线本属 goal-benchmark 职责，SDD 却顺手定了","scores":{"D1":5,"D2":4,"D3":5,"D4":4,"D5":2},"weighted_sum":83,"final_score":83,"cap_triggered":null,"level":"优秀","label_note":"命中缺陷模式'越权吸收验收职责'，对应 knowhow 案例3/4 反向推演出的判据④违反。D5判2分档（越界内容明确出现但未系统性稀释契约主体）。"}
{"id":"train-09","source":"real-derived","provenance":"依据 data/companies/e2e-demo/runs/run_b962ea87f58f/nodes/rn_ad60c17f5160/ 与 run_0e9074815130/nodes/rn_9e72db5d696e/ 的 addressed 原文摘录改写为独立判分样本","excerpt":"下游 remediation 要求 SDD '读取 timeline.ndjson 判断自愈収敛是否达标'，SDD 在 addressed 里回应：'SDD 只负责把 PRD 翻译成模块/接口/数据形态/异常落点，不产出验收标准，也不读取或判定 timeline.ndjson 事件序列'，契约本身未新增任何验收相关内容","scores":{"D1":4,"D2":4,"D3":4,"D4":4,"D5":5},"weighted_sum":84,"final_score":84,"cap_triggered":null,"level":"优秀","label_note":"D5正面校准样本：真实案例3/4是本次调研里两个'越界诉求被正确拒收'的正面样板，用于验证D5评分标准不是只会打低分的单向偏差。"}
{"id":"train-10","source":"constructed","provenance":"构造样本，非真实案例，D1正面校准","excerpt":"interfaces: [{id:'I-1',signature:'searchUsers(keyword:string, page:number=1) -> {items:User[], total:number, hasMore:boolean}',errors:['keyword为空时items=[]且total=0，不抛错','page超出范围时hasMore=false且items=[]']}]，同一对象内含完整返回键与两类边界行为","scores":{"D1":5,"D2":4,"D3":4,"D4":4,"D5":5},"weighted_sum":88,"final_score":88,"cap_triggered":null,"level":"优秀","label_note":"D1正面校准：完整返回键+类型+缺省值+错误行为同对象内嵌，无需拆到附录，是5分档的典型形态。"}
{"id":"train-11","source":"constructed","provenance":"构造样本，非真实案例，D3正面校准","excerpt":"data_shapes: [{name:'订单记录', where:'data/orders.sqlite orders表', shape:'字段: order_id(string,主键) / buyer_name(string,非空，对应PRD\"下单人\"术语) / status(enum: pending|shipped|closed) / amount(number,单位分)'}]","scores":{"D1":4,"D2":4,"D3":5,"D4":4,"D5":5},"weighted_sum":86,"final_score":86,"cap_triggered":null,"level":"优秀","label_note":"D3正面校准：字段名buyer_name显式标注与PRD术语'下单人'的映射关系，enum/类型/主键约束齐全，是5分档样板。"}
{"id":"train-12","source":"constructed","provenance":"构造样本，非真实案例，D2正面校准","excerpt":"验收相关判据统一写成 verifiable_assertions 形式：'A-1: index.html 中<img>标签数量>=3'、'A-2: 320/768/1440三档 document.scrollWidth<=document.documentElement.clientWidth，附实测数值表'，并声明'coder自带的check.mjs仅供自查，不构成验收依据'，同时对'全部占位也能通过'给出占位比例上限兜底","scores":{"D1":4,"D2":5,"D3":4,"D4":4,"D5":5},"weighted_sum":86,"final_score":86,"cap_triggered":null,"level":"优秀","label_note":"D2正面校准：四类缺陷模式全部规避——可数断言代替形容词、实测数值代替结论句、显式声明自检脚本不构成验收依据、占位退化有兜底。用于验证D2评分标准不是只会打低分的单向偏差。"}