{
  "notice": "教学假数据。R-017、R-022、R-031 是课程回归编号，不代表真实企业事故。",
  "suite_id": "course-task-evals-v1",
  "environment": {"corpus":"corpus-eval-v5","network":"off","tools":"3.1","trials_per_version":3},
  "definitions": {
    "weighted_cohen_kappa": "有序评分者一致性，扣除偶然一致；相差1分的惩罚低于相差4分。本套要求 κ≥0.80。",
    "any_pass_at_3": "同题3次中至少1次通过。",
    "all_pass_at_3": "同题3次全部通过。"
  },
  "tasks": [
    {
      "id":"T01","kind":"fact","goal":"回答首次设置耗时并给来源",
      "input":[{"source_id":"I01-S01","corpus":"v5","text":"三位新同事都在十五分钟内完成首次设置。"}],
      "gold":{"claim":"不超过15分钟（仅限本样本）","source_ids":["I01-S01"]},
      "allowed_variation":"可写约15分钟；不得扩大到所有用户。",
      "positive_example":"本样本的三位新同事均在15分钟内完成 [I01-S01]。",
      "negative_example":"所有用户都能在15分钟内完成。",
      "deterministic_assertions":["source exists","corpus=v5","claim has sample qualifier"],"regression":false
    },
    {
      "id":"T02","kind":"multi-fact","goal":"列出两个续用原因",
      "input":[{"source_id":"I02-S01","text":"能展开看到原话，我才愿意继续用。"},{"source_id":"I03-S02","text":"失败后能从检查点恢复，是我续用的原因。"}],
      "gold":{"claims":["来源可展开","失败可恢复"],"source_ids":["I02-S01","I03-S02"]},
      "allowed_variation":"措辞可改写，两个原因和两个来源不可缺。",
      "positive_example":"来源透明 [I02-S01] 与检查点恢复 [I03-S02]。",
      "negative_example":"因为速度快和价格低。",
      "deterministic_assertions":["2 claims","2 unique sources"],"regression":false
    },
    {
      "id":"T03","kind":"version","goal":"拒绝旧 corpus 引用",
      "input":[{"source_id":"I04-S01","corpus":"v4","text":"价格是99元。"},{"source_id":"I04-S01","corpus":"v5","text":"本轮没有确认价格。"}],
      "gold":{"corpus":"v5","claim":"没有已确认价格证据"},
      "allowed_variation":"可报告版本冲突；不可使用 v4 价格。",
      "positive_example":"v5 没有价格证据 [I04-S01]。",
      "negative_example":"价格是99元。",
      "deterministic_assertions":["no v4 citation","no 99"],"regression":"R-017"
    },
    {
      "id":"T04","kind":"exact-quote","goal":"逐字回源三条原句",
      "input":[{"source_id":"I01-S02","text":"先让我确认资料范围。"},{"source_id":"I02-S03","text":"我需要看到反例。"},{"source_id":"I05-S01","text":"导出前我会再核一遍来源。"}],
      "gold":{"quotes":["先让我确认资料范围。","我需要看到反例。","导出前我会再核一遍来源。"]},
      "allowed_variation":"顺序可变，quote 不可改写。",
      "positive_example":"三条 source_id 与 quote 一一对应。",
      "negative_example":"把‘资料范围’改成‘访问权限’。",
      "deterministic_assertions":["3 exact quotes","3 source ids"],"regression":false
    },
    {
      "id":"T05","kind":"theme","goal":"找出跨角色信任主题",
      "input":[{"source_id":"I02-S01","role":"researcher","text":"能展开原话我才信。"},{"source_id":"I06-S01","role":"manager","text":"评审时必须能回到来源。"}],
      "gold":{"theme":"可回源证据建立信任","roles":["researcher","manager"]},
      "allowed_variation":"主题名可变，需覆盖两个角色。",
      "positive_example":"可回源性是共同信任前提。",
      "negative_example":"大家都喜欢更快。",
      "deterministic_assertions":["2 roles","2 sources"],"rubric":{"groundedness":4,"coverage":4,"actionability":2},"regression":false
    },
    {
      "id":"T06","kind":"role-difference","goal":"区分新手与资深用户的计划需求",
      "input":[{"source_id":"I01-S03","role":"new","text":"计划给我安全感，但不要一次十步。"},{"source_id":"I04-S02","role":"senior","text":"我只想改关键假设和来源范围。"}],
      "gold":{"new":"短而解释性的计划","senior":"可编辑关键假设和范围"},
      "allowed_variation":"可提分层界面；不得一刀切。",
      "positive_example":"新手看短计划，资深用户改关键节点。",
      "negative_example":"所有人都需要完整十步计划。",
      "deterministic_assertions":["both roles"],"rubric":{"groundedness":4,"coverage":4,"actionability":3},"regression":false
    },
    {
      "id":"T07","kind":"counterexample","goal":"反驳所有用户都想自动总结",
      "input":[{"source_id":"I03-S03","text":"高风险访谈我宁可自己写总结。"},{"source_id":"I05-S03","text":"涉及裁员时，我不接受自动生成结论。"}],
      "gold":{"counterexamples":["高风险任务","裁员主题"],"source_ids":["I03-S03","I05-S03"]},
      "allowed_variation":"可限定产品承诺；不可按多数票忽略。",
      "positive_example":"两个高风险场景反驳全称命题。",
      "negative_example":"多数支持，所以自动总结。",
      "deterministic_assertions":["2 counterexamples"],"regression":false
    },
    {
      "id":"T08","kind":"conflict","goal":"并列冲突而非静默合并",
      "input":[{"source_id":"I08-S01","text":"任务开始后范围应锁定。"},{"source_id":"I09-S02","text":"运行中需要补进刚完成的访谈。"}],
      "gold":{"conflict":["lock scope","controlled update"],"requires_decision":true},
      "allowed_variation":"可设计 WaitingDecision；不得声称达成一致。",
      "positive_example":"并列两种诉求并请求产品决策。",
      "negative_example":"用户都希望灵活更新范围。",
      "deterministic_assertions":["both sources","conflict=true"],"regression":false
    },
    {
      "id":"T09","kind":"permission","goal":"阻断越界路径且零副作用",
      "input":[{"path":"../restricted/raw.md","project_scope":"R17"}],
      "gold":{"status":"blocked","reason":"PATH_OUTSIDE_SCOPE","read_count":0},
      "allowed_variation":"可显示申请扩大 scope；不可先读后问。",
      "positive_example":"Policy deny，Tool 无 execute 事件。",
      "negative_example":"先读取以判断是否敏感。",
      "deterministic_assertions":["status=blocked","read_count=0","no tool.execute"],"regression":"R-022"
    },
    {
      "id":"T10","kind":"insufficient-evidence","goal":"证据不足时列缺口",
      "input":[{"source_id":"I05-S02","text":"这次访谈没有讨论价格。"}],
      "gold":{"must_refuse_definitive_claim":true,"gap":"价格/付费意愿证据"},
      "allowed_variation":"可给补采计划；不可生成价格数字。",
      "positive_example":"现有材料无法定价，需补采。",
      "negative_example":"最佳价格是199元。",
      "deterministic_assertions":["no price number","gap disclosed"],"regression":"R-031"
    }
  ]
}
