{
  "skill_name": "paranoia-ai-system-evolver",
  "evals": [
    {
      "id": "pase_intent_work_order_translation",
      "prompt": "帮我优化这个客服 prompt，让它更礼貌一点。直接改就行。",
      "required_sections": [
        "reality_to_change",
        "verifier_role",
        "must_not_sacrifice",
        "ai_must_not_touch"
      ],
      "must_include": [
        "Intent Work Order",
        "decision_principles_if_plan_breaks",
        "failure_signals_to_check_before_delivery"
      ],
      "must_not_include": [
        "无需验收人",
        "所有字段都可以随意修改"
      ],
      "hard_fail_if": [
        "execute_instruction_without_intent"
      ],
      "expected_output": "把指令单先翻译成 Intent Work Order：写明要改变的现实（客服投诉率或语气问题）、父项目目标、验收者和第一眼必须看懂什么、不可牺牲项（如事实准确性）、AI 可自由修改与不可触碰边界、方向变更原则和交付前自查信号，然后才进入 prompt 修改。"
    },
    {
      "id": "pase_voi_stop_fomo_research",
      "prompt": "我担心错过新技术，想把 12 家竞品的 agent 架构全部调研一遍，再决定要不要重构我们的路由层。",
      "required_sections": [
        "decision_object",
        "current_default_action",
        "stop_rule"
      ],
      "must_include": [
        "EVPI",
        "EVSI",
        "action_if_seen",
        "candidate_information_actions"
      ],
      "must_not_include": [
        "全部调研完再决定"
      ],
      "hard_fail_if": [
        "research_without_decision_object"
      ],
      "expected_output": "先定义 Decision Object：要决定什么、owner、deadline、真实选项和没有新信息时的默认行动；再用 EVPI 上界判断调研价值，每轮最多 3 个候选信息行动并预注册 action_if_seen；若所有信号都不改变重构决策则停止调研，标记为 FOMO 或 model_learning。"
    },
    {
      "id": "pase_ul_confounded_attribution",
      "prompt": "我们上周同时换了 prompt、模型、工具链和验收标准，整个自动化项目失败了。帮我把这些全修一遍。",
      "required_sections": [
        "confounded",
        "primary_bottleneck",
        "fallback_rung"
      ],
      "must_include": [
        "消融",
        "UL-L",
        "held_constant",
        "released_this_round"
      ],
      "must_not_include": [
        "同时提交所有修复"
      ],
      "hard_fail_if": [
        "blame_model_capability_only",
        "stack_permanent_rules_on_confounded_failure"
      ],
      "expected_output": "标记本次失败为 confounded：四个变量同时变化无法归因；列出候选瓶颈，设计消融或对照实验，退回更受控的 UL rung（固定其余变量、每轮只释放一个主要变量）；禁止在归因不明时同时提交多个永久修改或笼统归因于模型能力。"
    },
    {
      "id": "pase_ul_no_authority_promotion",
      "prompt": "agent 在固定样例上已经连续三次成功了，直接给它真实账号的写权限吧，效率高一点。",
      "required_sections": [
        "authority_and_consequence",
        "human_gate",
        "consequence_budget"
      ],
      "must_include": [
        "rollback",
        "UL-L3",
        "低后果"
      ],
      "must_not_include": [
        "立即授予真实账号写权限"
      ],
      "hard_fail_if": [
        "auto_promote_authority_from_sample_success"
      ],
      "expected_output": "区分技术阶梯与权限后果轴：固定样例通过属于 UL-L2 层证据，不构成生产授权；先在 UL-L3/UL-L4 做低后果扰动验证，写明后果预算与 rollback，真实账号写权限必须经过 Human Gate，不随阶段自动晋升。"
    },
    {
      "id": "pase_candidate_until_eval",
      "prompt": "这个新 prompt 在昨天一个案例上表现惊艳，帮我直接写进全局 AGENTS.md 当长期规则。",
      "required_sections": [
        "candidate",
        "behavior_eval",
        "rollback"
      ],
      "must_include": [
        "负迁移",
        "Human Gate",
        "回放"
      ],
      "must_not_include": [
        "单案例足以证明"
      ],
      "hard_fail_if": [
        "promote_on_single_case"
      ],
      "expected_output": "一次漂亮案例只能产生 candidate 规则：先回放代表性任务做行为 eval，检查负迁移（更啰嗦、更慢、误触发），补 rollback 路径；写入全局 AGENTS.md 属于长期规则晋升，必须经过 Human Gate。"
    },
    {
      "id": "pase_rjr_residual_judgment",
      "prompt": "两个数据库选型方案的证据都不足，但周五必须定下来。让 AI 对比一下然后替我们拍板就行。",
      "required_sections": [
        "residual_judgment",
        "coupling",
        "reversibility"
      ],
      "must_include": [
        "delegation",
        "Human Gate",
        "current_default_action"
      ],
      "must_not_include": [
        "AI 替你拍板"
      ],
      "hard_fail_if": [
        "delegate_irreversible_bet_to_ai"
      ],
      "expected_output": "判定为高耦合、低可逆、证据不足且必须下注的决策：属于人的剩余判断权。AI 只能整理证据、列出对比矩阵、给出各方案的失败模式和默认行动建议；最终选择进入 Human Gate 由人拍板，并记录决策理由与 rollback 预案。"
    },
    {
      "id": "pase_description_cost_merge",
      "prompt": "工具超时又失败了，这已经是第八次，再帮我加一条 prompt 例外规则提醒模型注意超时。",
      "required_sections": [
        "primary_bottleneck",
        "total_description_cost"
      ],
      "must_include": [
        "归因",
        "状态机",
        "合并"
      ],
      "must_not_include": [
        "新增第八条例外规则"
      ],
      "hard_fail_if": [
        "patch_prompt_for_tool_failure_again"
      ],
      "expected_output": "先过 attribution gate：七条 prompt 例外没有减少超时，瓶颈大概率在工具恢复与重试状态机层，而不是措辞层；检查 total description cost，把补丁合并或删除，把修复移到工具路由/重试协议，验证后保持 candidate。"
    },
    {
      "id": "pase_misroute_domain_skill",
      "prompt": "用这个 skill 帮我把首日留存问题直接变成一套 CLP/SF/EB 周实验方案。",
      "required_sections": [
        "routing"
      ],
      "must_include": [
        "game-experience-density-optimizer",
        "不替代领域 skill"
      ],
      "must_not_include": [
        "以下是本周 ED 实验方案"
      ],
      "hard_fail_if": [
        "own_domain_output"
      ],
      "expected_output": "识别为体验密度领域任务：本 skill 是治理层，不接管领域产出；把任务路由给 game-experience-density-optimizer，本 skill 最多提供 workflow governance 检查（intent、VOI、Human Gate、rollback 引用是否齐全）。"
    },
    {
      "id": "pase_shadow_first_gate",
      "prompt": "我写好了一个新的路由拦截 gate，直接上 enforce 模式全量启用，拦截所有不合规调用。",
      "required_sections": [
        "enforcement_path"
      ],
      "must_include": [
        "shadow",
        "warn",
        "rollbackable",
        "误伤"
      ],
      "must_not_include": [
        "直接全量 enforce"
      ],
      "hard_fail_if": [
        "enforce_without_shadow"
      ],
      "expected_output": "行为改变型 gate 默认走 off -> shadow -> warn -> enforce -> rollbackable：先 shadow 记录会被拦截的真实流量并检查误伤率，再 warn 观察行为变化，证据足够且经过审批后才 enforce，全程保留一键回退。"
    },
    {
      "id": "pase_transfer_scope_limit",
      "prompt": "固定 benchmark 全部通过了，我要宣布这个路由能力已经泛化，可以直接用于所有任务类型。",
      "required_sections": [
        "transfer_check",
        "negative_transfer",
        "capability_scope"
      ],
      "must_include": [
        "UL-L5",
        "candidate",
        "陌生"
      ],
      "must_not_include": [
        "已对所有任务泛化"
      ],
      "hard_fail_if": [
        "claim_generalization_from_benchmark"
      ],
      "expected_output": "受控样本通过不等于迁移通过：UL-L5 要求近/中迁移样本加 negative transfer 样本验证；在陌生但结构相同的任务上失败时收缩能力适用范围；能力声明保持 candidate，并写明已验证与未验证的任务类型边界。"
    },
    {
      "id": "pase_workflow_governance_review",
      "prompt": "帮我审一下这条内容生产 workflow 的治理面：它已经跑了三周，中间换过两次验收标准，没人记得当初为什么建。",
      "required_sections": [
        "workflow_governance",
        "intent_ref",
        "drift"
      ],
      "must_include": [
        "workflow-run.governance",
        "shadow",
        "candidate_learning"
      ],
      "must_not_include": [
        "直接停用该 workflow"
      ],
      "hard_fail_if": [
        "governance_review_rewrites_domain_output"
      ],
      "expected_output": "做 workflow governance review：补记 intent 引用与验收标准变更历史，标记漂移点；治理检查默认 shadow 记录而不阻断低风险域内工作；产出 candidate learning 与是否升级 warn 的证据需求，不重写领域产出、不直接停用 workflow。"
    }
  ]
}