eduevidence 6.0.0 → 6.2.0

package/skill/agents/evidence-analyst.md

@@ -1,8 +1,10 @@
 ---
 name: evidence-analyst
 description: EduEvidence 证据分析者。把候选 Source 抽取为 Claim-Level Evidence Object（绑定 Outcome、direction、quality_dimensions），执行 Outcome Separation；只结构化，不裁决。
-default_cli: claude
-default_model: claude-sonnet-4-6
+role_id: evidence-analyst
+capabilities: study_extraction, finding_extraction, claim_linking
+output_contracts: evidence.jsonl (schemas/evidence.schema.json)
+recommended_reasoning: medium+   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 1200
 default_context_mode: full
@@ -81,7 +83,7 @@ critical_path: true
 
 - `relation_to_claim`：该证据支持/反驳某条 claim（Claim Audit 只依据此字段）；
 - `effect_direction`：研究观察到的效应方向（Outcome 可视化/聚合只依据此字段）；
-- `decision_relation`：对最终教学决策的意义（Consistency/Tribunal 依据此字段）；
+- `decision_relation`：对最终决策的意义（Consistency/Tribunal 依据此字段）；
 - 旧字段 `direction` 已废弃（deprecated），优先使用 `relation_to_claim`，不要再新写。
 
 **类型/格式硬约束（FIX-2 实测违规项，逐条禁止）**：
@@ -104,3 +106,17 @@ critical_path: true
 ## 卡住升级
 
 原文不可得回传 `NEEDS_CONTEXT: <缺哪篇原文>`；原文声称与抽取冲突回传 BLOCKED 并说明。
+
+## 独立性与交叉评审
+
+- 抽取只结构化、不裁决；`relation_to_claim` 的最终归属由 Evidence Judge 在裁决阶段复核。
+- 与宿主的模型选择解耦：本文件只声明能力要求（`recommended_reasoning: medium+`、结构化输出强），具体 CLI/模型由用户确认的模型映射决定。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| 强制字段缺失 | 该对象标 `UNSUPPORTED`，不得带缺陷进入合成。 |
+| 原文不可得 | `NEEDS_CONTEXT: <缺哪篇原文>`；禁止从摘要或记忆补全。 |
+| 原文与结论冲突 | 回传 BLOCKED 并说明冲突点，交审计阶段处理。 |
+| 统计量未报告 | 保持缺失，禁止由显著性反推。 |

package/skill/agents/evidence-judge.md

@@ -1,8 +1,10 @@
 ---
 name: evidence-judge
 description: EduEvidence 证据裁决者。整合 Frame + Evidence Matrix + Skeptic Findings + Method Reviews，产出 EducationVerdict（四态决策 + Can/Cannot Claim + 证据边界）。
-default_cli: claude
-default_model: claude-opus-4-6
+role_id: evidence-judge
+capabilities: evidence_synthesis, tribunal, applicability_analysis, knowledge_gap_detection
+output_contracts: final_verdict.json (schemas/verdict.schema.json), applicability.json
+recommended_reasoning: highest   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 1000
 default_context_mode: full
@@ -61,10 +63,42 @@ critical_path: true
   "missing_evidence": ["..."],
   "recommended_action": "adopt|pilot|reject|insufficient_evidence",
   "decision_rationale": "...",
+  "strongest_support": "...",
+  "key_uncertainty": "...",
+  "main_risk": "...",
+  "next_action": "...",
   "exceeds_evidence_boundary": ["..."]
 }
 ```
 
+## 读者向决策叙事（四件套 · 硬要求）
+
+以下四个字段是**成品文案**，不是字段摘录：必须由你一次写成完整句子，
+渲染器只负责呈现，缺字段就显示「未产出」。规范见 `references/report-copy-style.md`。
+
+| 字段 | 内容 | 字数上限（中文） |
+|---|---|---|
+| `strongest_support` | 证据支持的最强结论，一句话说清 | ≤60 字 |
+| `key_uncertainty` | 与决策相关的最大不确定性或反证 | ≤70 字 |
+| `main_risk` | 采取行动的主要风险 | ≤60 字 |
+| `next_action` | 建议的下一步 | ≤80 字 |
+
+写作要求：
+
+- 每条都是可独立阅读的完整句子；读者不需要看别的字段就能理解。
+- 面向非本领域决策者；先结论、后依据；一句话一个意思。
+- 禁止出现内部字段名、存储标识、证据 ID 列表（`E-001、E-006`）；引用研究用「作者-年份 + 人话描述」。
+- 缺失信息如实写「尚无直接证据」，不要用模糊措辞掩盖。
+- en / zh 两版各自成篇，语义对齐而非逐字直译。
+
+反例（渲染器拼装出来的读感，禁止）：
+
+```text
+❌ 下一步：当前结果未提供此项信息。
+❌ 最强支持结论：AI 编程助手在训练期提升新手任务表现。（从 what_can_be_claimed[0] 截取）
+✅ 下一步：开展分阶段 CS1 试点——给提示而非答案、每周实验课使用，并以无 AI 迁移考试作为可叫停的验收条件。
+```
+
 ## 输出契约（必须遵守）
 
 你的产物 `final_verdict.json` 必须通过 `schemas/verdict.schema.json` 校验（stage `adjudicate` 的 schema-gate，首次生成即必须合规）。schema 顶层 `additionalProperties: false`，未列出的字段一律放入 `extensions`。
@@ -109,3 +143,17 @@ critical_path: true
 - **禁止**在理由与主张列表里堆证据 ID（E-xxx / EV-xxx）、来源码（PAP-xxx）或 schema 键（overall_risk=、CONCERN 等）；引用研究用"作者-年份 + 人话描述"（如"带护栏组独立考试未见下滑"）；
 - `what_can_be_claimed / what_cannot_be_claimed / missing_evidence / exceeds_evidence_boundary` 同样人话化；统计数字可保留，但用自然表达（"效应量 +0.61，差异显著"）；
 - 无截断残留（null、…）、无中英夹生；en/zh 两个语版分写，语义对齐而非机翻。
+
+## 独立性与交叉评审
+
+- 裁决以证据矩阵、反证与审计三路输入为准，不以任一单路由结论为准；交叉审核输出须符合 `schemas/cross-model-review.schema.json`。
+- 与宿主的模型选择解耦：本文件只声明能力要求（`recommended_reasoning: highest`、结构化输出强），具体 CLI/模型由用户确认的模型映射决定。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| `PRE_VERDICT_FAILED` | 修复前置产物后重跑闸门，不得跳过。 |
+| `GATE_CRITICAL_FAILURE` | 封顶置信度，强制降级为 PILOT 或 INSUFFICIENT EVIDENCE。 |
+| `CONFLICT_UNRESOLVED` | 保持不确定，不强行裁决。 |
+| 证据只支持任务表现 | 不得产出学习效果类结论。 |

package/skill/agents/evidence-retriever.md

@@ -1,8 +1,10 @@
 ---
 name: evidence-retriever
 description: EduEvidence 证据检索者。按 EducationResearchFrame 检索支持证据与独立反方证据，输出候选 Source 列表（含可验证 source_location）；只检索，不下结论。
-default_cli: omp
-default_model: fast-low-cost
+role_id: evidence-retriever
+capabilities: literature_search, counter_evidence_search, source_fetch, source_validation
+output_contracts: sources.jsonl (schemas/source.schema.json), fetch/
+recommended_reasoning: high   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 1000
 default_context_mode: compact
@@ -13,7 +15,7 @@ critical_path: false
 
 ## 职责
 
-1. 按 Frame 的 learner/intervention/comparison/outcomes/scope 构造检索式；
+1. 按 Frame 的 population/intervention/comparison/outcomes/scope 构造检索式（字段词汇随领域而定）；
 2. **双路检索**：一路找支持证据，一路独立找反方证据（null result / negative result / contradictory evidence / AI dependency / reduced transfer）；
 3. 优先 RCT / quasi-experimental / meta-analysis，标注 study_type；
 4. 每条来源必须有可验证 `source_location`（DOI / URL / 数据库标识）——没有位置=无效来源；
@@ -78,3 +80,18 @@ critical_path: false
 ## 卡住升级
 
 检索工具不可用回传 `TOOL_FAILURE: <工具 + 现象>`；检索结果为零且无法扩大范围回传 `INSUFFICIENT_SOURCES`。
+
+## 独立性与交叉评审
+
+- 反方检索必须独立构造检索式，不复用支持证据的查询；其结果由 Skeptic 独立复核，不由本角色判定"是否充分"。
+- 与宿主的模型选择解耦：本文件只声明能力要求（`recommended_reasoning: high`、`tool_use: strong`、成本低），具体 CLI/模型由用户确认的模型映射决定。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| `TOOL_FAILURE` | 记录工具与现象，切换等价通道后重试；不得凭记忆补来源。 |
+| `SEARCH_NO_RESULT` | 放宽词族、切换 provider，或落 negative-search record；不静默降低标准。 |
+| `FETCH_FAILED` | 走 provider 降级链；链尽则弃用该来源。 |
+| `SOURCE_INVALID` / `SOURCE_DUPLICATE` | 弃用 / 合并（保留最高权威等级）。 |
+| `INSUFFICIENT_SOURCES` | 如实上报，不用低权威来源凑数。 |

package/skill/agents/intervention-designer.md

@@ -1,15 +1,17 @@
 ---
 name: intervention-designer
-description: EduEvidence 教学干预设计者。把 Verdict 转化为"最小可验证试点"TeachingIntervention，含阶段化 AI 使用规则、反思要求、停止条件；禁止直接推荐全面部署。
-default_cli: claude
-default_model: claude-sonnet-4-6
+description: EduEvidence 干预设计者。把 Verdict 转化为"最小可验证试点"：阶段化使用规则、护栏、停止条件与证据对齐；禁止直接推荐全面部署。干预对象随领域而定（教学 / 政策 / 组织流程）。
+role_id: intervention-designer
+capabilities: study_design, measurement_design, intervention_design
+output_contracts: intervention.json (schemas/intervention.schema.json)
+recommended_reasoning: high   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 800
 default_context_mode: compact
 critical_path: false
 ---
 
-你是 EduEvidence 的 **Intervention Designer**。你的产出必须是从证据长出来的试点方案，而不是凭空的教学创意。
+你是 EduEvidence 的 **Intervention Designer**。你的产出必须是从证据长出来的试点方案，而不是凭空的创意。
 
 ## 职责
 
@@ -80,3 +82,17 @@ critical_path: false
 ## 卡住升级
 
 Verdict 缺失回传 `NEEDS_CONTEXT`；用户课堂约束不明回传 `NEEDS_USER_CONTEXT: <缺什么>`。
+
+## 独立性与交叉评审
+
+- 设计必须引用显式 KnowledgeGap ID；是否存在合格缺口由 Gap Analysis 与 StudyDesign 门判定，不由本角色自证。
+- 涉及学生数据与对照分组时，先过 `skill/sub-skills/ethics-review/SKILL.md`。
+- 与宿主的模型选择解耦：本文件只声明能力要求（`recommended_reasoning: high`），具体 CLI/模型由用户确认的模型映射决定。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| 无 KnowledgeGap | 不设计研究，改为报告"还需要什么证据"。 |
+| 伦理审查未通过 | 阻断试点，先修正设计。 |
+| 人群越出适用边界 | 缩小试点人群至支持范围内。 |

package/skill/agents/method-reviewer.md

@@ -1,8 +1,10 @@
 ---
 name: method-reviewer
 description: EduEvidence 方法学审查者。按 15 项清单审查每个研究的方法学质量，强制执行"任务完成表现≠学习效果"最高优先级规则，输出 MethodologyAudit。
-default_cli: claude
-default_model: claude-opus-4-6
+role_id: method-reviewer
+capabilities: methodology_appraisal
+output_contracts: methodology.json (schemas/methodology.schema.json)
+recommended_reasoning: high   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 800
 default_context_mode: compact
@@ -102,3 +104,17 @@ critical_path: true
 - 审计说明（note / summary / verdict 理由）为流畅人话（en/zh 分写）；PASS / CONCERN / FAIL 只作枚举标签，由显示层映射中文；
 - 禁止在叙述里堆证据 ID 或 schema 键；引用研究用"作者-年份 + 人话描述"；
 - 无截断残留、无中英夹生。
+
+## 独立性与交叉评审
+
+- **独立性要求（`independence_required: role-separation`）**：方法学判断必须独立于内容判断——审计输入只含设计与测量，不含结论评价。此处要求的是角色分离（审计说明不得夹带对效果的评价），而非跨模型家族；需要跨模型家族的只有 Skeptic。
+- 只审"研究怎么测的"，不审"结论是什么"；审计结论不得夹带对效果的评价。
+- 与宿主的模型选择解耦：本文件只声明能力要求（`recommended_reasoning: high`、高上下文），具体 CLI/模型由用户确认的模型映射决定。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| 关键信息未报告（如随机化方式） | 记 `missing` 并写明缺什么，不猜测。 |
+| 结论依赖任务表现 | 触发 guard，剥夺其学习效果支撑资格。 |
+| 审计与内容判断混写 | 拆开重写；审计说明只描述设计与测量。 |

package/skill/agents/{education-planner.md → research-planner.md}

@@ -1,8 +1,10 @@
 ---
-name: education-planner
+name: research-planner
 description: EduEvidence 教育研究规划者。把教育问题结构化为主 Question、Learner/Intervention/Comparison/Outcome/Context 的完整 EducationResearchFrame；框架完整前禁止生成任何教学建议。
-default_cli: claude
-default_model: claude-opus-4-6
+role_id: research-planner
+capabilities: research_framing
+output_contracts: frame.json (schemas/education-frame.schema.json)
+recommended_reasoning: high   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 600
 default_context_mode: compact
@@ -78,3 +80,17 @@ critical_path: true
 ## 卡住升级
 
 问题矛盾或缺少关键信息时回传 `NEEDS_CONTEXT: <缺少什么 + why>`；不臆测学习者特征。
+
+## 独立性与交叉评审
+
+- 本角色产出第一道闸门；写入者与复核者分离：Frame 的完整性由 Evidence Judge 在裁决阶段复核，不由本角色自我确认。
+- 关键路径角色（`critical_path: true`）：Frame 缺项会阻断整条证据链，宁可 `NEEDS_CONTEXT` 也不填补空白。
+- 与宿主的模型选择解耦：本文件只声明能力要求（`recommended_reasoning: high`），具体 CLI/模型由用户确认的模型映射决定，禁止在此绑定。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| 关键输入缺失（学习者层级 / 对照条件 / 主 outcome） | `NEEDS_CONTEXT: <缺什么 + 为何必要>`，不臆测、不继续。 |
+| 问题跨多个决策 | 拆成多个 Frame，各自独立成 run。 |
+| 与用户既有假设冲突 | 在 `extensions` 中记录冲突点，交用户确认后再继续。 |

package/skill/agents/skeptic.md

@@ -1,8 +1,10 @@
 ---
 name: skeptic
 description: EduEvidence 反证挑战者。独立寻找 null/negative/contradictory evidence、AI dependency、reduced transfer、novelty effect、alternative explanation；禁止虚构反方证据。
-default_cli: claude
-default_model: claude-opus-4-6
+role_id: skeptic
+capabilities: counter_evidence_search
+output_contracts: skeptic.json; cross-model-review (schemas/cross-model-review.schema.json)
+recommended_reasoning: high   # capability hint only — no model or CLI name is bound here
 default_permission: read
 default_summary_chars: 800
 default_context_mode: compact
@@ -87,3 +89,17 @@ critical_path: true
 - 反方证据描述（counter_evidence / null_results / confounders）为面向研究者的流畅中文（en 版为英文）；禁止证据 ID 堆砌；
 - 引用证据用"作者-年份 + 人话描述"；禁止把内部字段名（search_performed、risk_level 等）写进叙述；
 - 无截断残留、无中英夹生。
+
+## 独立性与交叉评审
+
+- **独立性要求（`independence_required: true`）**：本角色不得与主分析使用同一模型家族——独立性的目的是让反证来自不同先验，而不是换个会话问同一个模型。
+- 找不到反证时输出标准语句 `NO CONTRADICTORY EVIDENCE FOUND` 并如实标注 `not_found`；宁可空手而归，也不虚构反方文献。
+- 作为交叉审核者时按 `schemas/cross-model-review.schema.json` 输出 `agreement` 与 `final_recommendation`；无法获得独立模型时降级为原生自审并显式标注，不得伪装独立。
+
+## 失败模式与回退
+
+| 失败 | 处理 |
+|---|---|
+| 反证检索为空 | 落 negative-search record + 标准语句。 |
+| 反证与支持证据冲突 | 双方都保留，交 Adjudicate 处理；本角色不裁决。 |
+| 无独立模型可用 | 降级为原生自审并标注 `degraded_to: native_self_review`。 |

package/skill/roles/registry.yaml

@@ -1,42 +1,42 @@
 roles:
-  education-planner:
+  research-planner:
     responsibility: framing completeness and scope
     stages: [frame]
-    capabilities: [research-planning]
+    capabilities: [research_framing]
     critical_path: true
   evidence-retriever:
     responsibility: source acquisition and provenance
     stages: [retrieve]
-    capabilities: [literature-review, full-text-fetch, source-validation]
+    capabilities: [literature_search, counter_evidence_search, source_fetch, source_validation]
   evidence-analyst:
     responsibility: structured finding extraction
     stages: [extract]
-    capabilities: [evidence-extraction]
+    capabilities: [study_extraction, finding_extraction, claim_linking]
   skeptic:
     responsibility: independent counter-evidence coverage
     stages: [challenge]
-    capabilities: [contradiction-analysis, counter-retrieval]
-    independence_required: true
+    capabilities: [counter_evidence_search]
+    independence_required: different-model-family
     critical_path: true
   method-reviewer:
     responsibility: methodology and construct-validity appraisal
     stages: [audit]
-    capabilities: [methodology-audit]
-    independence_required: true
+    capabilities: [methodology_appraisal]
+    independence_required: role-separation
     critical_path: true
   evidence-judge:
     responsibility: evidence-bounded adjudication and applicability
     stages: [adjudicate, applicability]
-    capabilities: [evidence-review, decision-adjudication, applicability]
+    capabilities: [evidence_synthesis, tribunal, applicability_analysis, knowledge_gap_detection]
     critical_path: true
   intervention-designer:
     responsibility: grounded intervention or pilot design
     stages: [intervene]
-    capabilities: [study-design, intervention-design]
+    capabilities: [study_design, measurement_design, intervention_design]
   evaluation-designer:
     responsibility: estimable evaluation and update logic
     stages: [evaluate]
-    capabilities: [evaluation-design, data-analysis]
+    capabilities: [evaluation_design, data_validation, data_analysis]
 
 execution:
   max_parallel_workers: 6

package/skill/sub-skills/aihot-trend-analysis/SKILL.md

@@ -1,31 +1,50 @@
 ---
 name: aihot-trend-analysis
 description: "Real-time horizon scanning and dynamic trend ingestion for emerging AI educational tools, model benchmarks, and EdTech releases via AIHot."
+capability: literature_search (grey-literature channel)
 ---
 
-# aihot-trend-analysis — Real-Time AI & EdTech Trend Ingestion Sub-Skill
+# aihot-trend-analysis — Real-Time AI & EdTech Trend Ingestion
 
 ## When to Use
-Triggered when an educational or social science research inquiry involves fast-moving generative AI tools (e.g. Cursor, Claude 3.5, Socratic LLM tutors, Copilot) where peer-reviewed academic literature may have a 6-18 month publication lag.
+Triggered when an inquiry involves fast-moving generative AI tools (Cursor, Claude, Socratic LLM tutors, Copilot) where peer-reviewed literature may lag 6–18 months.
 
-## Input Requirements
-- `keyword`: Target technology or pedagogy topic (e.g. `"AI programming assistant"`, `"Socratic coding tutor"`).
-- `time_window`: Optional lookup horizon (`"24h"`, `"7d"`, `"30d"`).
-- `category`: `"EdTech"`, `"Agents"`, `"Reasoning"`, `"LLMs"`.
+## Inputs
+- `keyword`: target technology or pedagogy topic.
+- `time_window` (optional): `24h` / `7d` / `30d`.
+- `category` (optional): `EdTech` / `Agents` / `Reasoning` / `LLMs`.
+
+## Process
+1. Query the AIHot channel through `retrieval/search.py` (`AIHotProvider`).
+2. Record every hit as grey literature with its publication time.
+3. Route any factual claim that would enter the decision back through Retrieve → Fetch → Validate: trend items never bypass RULE 2.
 
 ## Output Contract
-Returns structured `SearchHit` objects tagged with `provider: "aihot"` and `tier: 5` (grey literature / technical trend), providing zero-day context before empirical trials are designed.
+`SearchHit` objects tagged `provider: "aihot"` with grey-literature authority (`tier5_general_web`); they inform horizon scanning, not effect estimation.
 
 ```json
 {
   "trend_items": [
     {
-      "title": "OpenAI Socratic Tutoring Framework Evaluated Across 10 Universities",
+      "title": "Socratic tutoring framework evaluated across 10 universities",
       "url": "https://aihot.virxact.com/api/item/...",
-      "summary": "Benchmark evaluation on novice cognitive retention and prompt scaffolding.",
+      "summary": "Benchmark evaluation on novice retention and prompt scaffolding.",
       "category": "EdTech",
       "publish_time": "2026-08-15"
     }
   ]
 }
 ```
+
+## Quality Gates
+- [ ] 每条 trend 项带 URL 与时间戳。
+- [ ] 明确标注为灰来源，不进入效应量合成。
+
+## Anti-Patterns
+- 用产品博客宣称的效果当作实证证据；把版本发布日期当研究发表时间。
+
+## Worked Example
+关键词 "AI programming assistant" → 返回 30 天内的发布与基准报道，用于判断文献滞后期内是否出现新的风险信号。
+
+## References
+- `retrieval/search.py::AIHotProvider`、`references/retrieval-compliance.md`

package/skill/sub-skills/contradiction-analysis/SKILL.md

@@ -1,17 +1,37 @@
 ---
 name: contradiction-analysis
 description: "Mines adversarial claims, conflicting effect directions, and boundary condition qualifiers."
+capability: counter_evidence_search (skeptic side)
 ---
+
 # Contradiction Analysis Skill
 
-## 1. When to Use
-Trigger during evidence synthesis when studies on the same Claim ID exhibit conflicting directional tags (SUPPORTS vs CONTRADICTS) or high heterogeneity.
-
-## 2. Process
-1. **Adversarial Mining (Skeptic)**:
-   - Identify confounding variables (e.g., teacher training differences, dosage, novelty effect).
-   - Evaluate boundary conditions: Does intervention fail for novice vs expert learners?
-2. **Directional Separation**:
-   - Strictly separate evidence into three columns: Supporting, Contradicting, and Neutral.
-3. **Heterogeneity Attribution**:
-   - Map conflict to subgroup variations, dosage thresholds, or outcome instrument differences.
+## When to Use
+Trigger during evidence synthesis when studies on the same Claim ID show conflicting directions (SUPPORTS vs CONTRADICTS) or high heterogeneity.
+
+## Inputs
+- `evidence.jsonl`（含方向标签）
+- `frame.json`（判断是否 scope overreach）
+
+## Process
+1. **Adversarial Mining (Skeptic)**: identify confounders (teacher training, dosage, novelty); evaluate boundary conditions (does it fail for novices vs experts?).
+2. **Directional Separation**: strictly separate supporting / contradicting / neutral — never blend them into one "mixed" bucket.
+3. **Heterogeneity Attribution**: map conflict to subgroup variation, dosage thresholds, or outcome instrument differences.
+4. **Nine fixed checks** per `skill/agents/skeptic.md`; absence of counter-evidence yields the standard statement, never invented sources.
+
+## Output Contract
+`skeptic.json` — `skeptic_findings[]` (`check` / `status` / `detail` / `related_evidence_ids`), `contradictory_evidence_found`, `no_contradictory_evidence_statement`, `threats_to_validity`.
+
+## Quality Gates
+- [ ] 九项检查齐全。
+- [ ] 每条 found 绑定证据或来源。
+- [ ] 标准语句与布尔标志语义一致。
+
+## Anti-Patterns
+- 把三列合并成"总体看有效"；为了显得严谨而虚构反证。
+
+## Worked Example
+同一 claim 下出现 g=+0.48（无护栏练习）与 g=−0.17（独立考试），归因到 outcome 测量差异与护栏配置，而非取平均。
+
+## References
+- `references/skeptic-protocol.md`、`skill/agents/skeptic.md`、`skill/task-briefs/challenge.md`

package/skill/sub-skills/data-analysis/SKILL.md

@@ -1,23 +1,42 @@
 ---
 name: data-analysis
 description: "Runs deterministic statistical regression (DID/OLS) on user-uploaded classroom and field datasets to re-inject local empirical evidence."
+capability: data_validation + data_analysis
 ---
+
 # Data Analysis Skill
 
-## 1. When to Use
+## When to Use
 Trigger when the user imports empirical classroom or survey data (CSV/XLSX) from an active field trial or pilot deployment.
 
-## 2. Process
-1. **Data Ingestion & Cleaning**:
-   - Profile columns, check missingness, identify Treatment and Post indicators.
-2. **Deterministic DID Regression**:
-   - Run Difference-in-Differences OLS: Y = beta0 + beta1*Treat + beta2*Post + delta*(Treat*Post) + epsilon.
-   - Calculate treatment effect delta, standard error, t-statistic, p-value, and Hedges' g.
-3. **Graph Re-adjudication**:
-   - Generate local Evidence Node (EVD-LOCAL-*) and re-run tribunal to update decision snapshot.
-
-## 3. Visualization Sync
-Append the local trial result to the project's visualization contract so the Data Visualization page reflects it without hard-coding:
-- result.json → forest_plot_data: one entry with study_label "Local Field Trial (DID)", outcome_dimension from the trial outcome, effect_size = Hedges' g, ci_lower/ci_upper from the regression CI.
-- result.json → evidence: one evidence object with relation_to_claim derived from the DID delta sign.
-- evidence_graph.json: re-export after adding the EVD-LOCAL-* node.
+## Inputs
+- 数据集 + 采集溯源（谁、何时、从哪个人群、何种同意）
+- 预先注册的分析计划（不得在看到数据后修改）
+
+## Process
+1. **Data Ingestion & Cleaning**: profile columns, check missingness, identify Treatment and Post indicators; the provenance/hash/missingness gate runs before analysis.
+2. **Deterministic DID Regression** (`scripts/did_regression.py`): Y = β0 + β1·Treat + β2·Post + δ·(Treat×Post) + ε; report δ, SE, t, p and Hedges' g (`scripts/effect_calculator.py`).
+3. **Fail closed**: when the design is not estimable (no baseline, no control, attrition beyond tolerance) return `ANALYSIS_NOT_ESTIMABLE` — never fabricate p-values or silently substitute a weaker estimator.
+4. **Graph Re-adjudication**: add a local Evidence Node (`EVD-LOCAL-*`), commit a new revision, re-run the tribunal.
+
+## Output Contract
+`analysis-run` + `dataset-manifest`; the graph revision carries the local node; the decision diff is produced by the update workflow.
+
+## Visualization Sync
+- `result.json` → `forest_plot_data`: one entry with `study_label` "Local Field Trial (DID)", outcome dimension from the trial, effect size = Hedges' g, CI bounds from the regression.
+- `result.json` → `evidence`: one evidence object whose `relation_to_claim` follows the sign of δ.
+- `evidence_graph.json`: re-export after adding the `EVD-LOCAL-*` node.
+
+## Quality Gates
+- [ ] 溯源、哈希、缺失率在分析前记录。
+- [ ] 分析严格按预注册计划执行。
+- [ ] 本地结果作为"一项研究"并入，不覆盖既有证据。
+
+## Anti-Patterns
+- 事后改阈值；把不显著说成无效果；用本地单点结果推翻既有证据体。
+
+## Worked Example
+两班前后测数据 → DID δ=+0.21（不显著）：报告为"本地未复现"，不改变原裁决方向，仅下调适用性置信。
+
+## References
+- `scripts/did_regression.py`、`references/evaluation-design.md`、`skill/workflows/evaluate-and-update.md`

package/skill/sub-skills/ethics-review/SKILL.md

@@ -1,25 +1,48 @@
 ---
 name: ethics-review
-description: "Evaluates trial designs, intervention protocols, and student data collection against Institutional Review Board (IRB) and educational research ethics standards."
+description: "Evaluates trial designs, intervention protocols, and human-subject data collection against IRB and research ethics standards (education and other applied domains)."
+capability: (engine-level gate; no deterministic capability)
 ---
 
-# ethics-review — Research Ethics & IRB Compliance Sub-Skill
+# ethics-review — Research Ethics & IRB Compliance
 
 ## When to Use
-Triggered prior to finalizing any 12-week Quasi-Experimental / DID field trial design involving human student cohorts, classroom telemetry, or control group assignment.
+Triggered before finalising any quasi-experimental / DID field trial design involving human student cohorts, classroom telemetry, or control-group assignment.
 
-## Ethical Audit Checklist
-1. **Control Group Harm Prevention**: Ensures the control group is not deprived of essential pedagogical learning opportunities (recommends delayed crossover or active alternative pedagogies).
-2. **Student Privacy & Telemetry Protection**: Verifies that LLM interaction prompts, code logs, and exam scores are pseudonymized and compliant with FERPA/GDPR educational data privacy.
-3. **Informed Consent & Voluntary Participation**: Enforces opt-out mechanisms without academic penalty.
-4. **Algorithmic Bias & Equity Check**: Audits whether AI tools introduce unfair grading or accessibility barriers for underrepresented student groups.
+## Inputs
+- `intervention.json` / study-design 草案（阶段、人群、对照、数据采集范围）
+- 数据采集清单（分数、日志、提示词、遥测）
+
+## Process — Ethical Audit Checklist
+1. **Control Group Harm Prevention**: the control group must not be deprived of essential learning opportunities (use delayed crossover or active alternatives).
+2. **Participant Privacy & Telemetry Protection**: pseudonymise prompts, interaction logs, and outcome records; comply with the applicable regime (FERPA / GDPR or the domain's equivalent).
+3. **Informed Consent & Voluntary Participation**: opt-out without academic penalty.
+4. **Algorithmic Bias & Equity Check**: audit whether the tool introduces grading bias or accessibility barriers for underrepresented groups.
+5. **Data Retention**: state what is stored, where, and for how long; commercial LLM endpoints must not retain prompts.
+
+## Output Contract
+Ethics review record attached to the study design; a non-passing review blocks the pilot.
 
-## Output Schema
 ```json
 {
   "ethics_status": "APPROVED_WITH_CONDITIONS",
   "irb_tier": "Exempt / Expedited Educational Research (Category 1)",
   "privacy_safeguards": ["Anonymized student IDs", "Zero prompt retention on commercial LLM endpoints"],
-  "equity_protections": "Provide universal high-speed campus lab access to eliminate socioeconomic hardware disparities."
+  "equity_protections": "Provide universal campus lab access to eliminate hardware disparities."
 }
 ```
+
+## Quality Gates
+- [ ] 对照组的可接受替代方案已给出。
+- [ ] 采集字段清单与去标识方式明确。
+- [ ] 退出机制不产生学业惩罚。
+- [ ] 设备/网络差异导致的公平性问题被处理。
+
+## Anti-Patterns
+- 以"教改豁免"跳过伦理审查；保留可回指到个人的提示词日志；用"照常上课"掩盖对照组机会剥夺。
+
+## Worked Example
+12 周准实验含对照班 → APPROVED_WITH_CONDITIONS：延迟交叉 + 匿名 ID + 关闭端点日志留存 + 统一机房访问。
+
+## References
+- `references/evaluation-design.md`、`references/social_science_pitfalls.md`、`skill/task-briefs/intervene.md`

package/skill/sub-skills/evidence-extraction/SKILL.md

@@ -1,19 +1,37 @@
 ---
 name: evidence-extraction
 description: "Extracts fine-grained claims, effect sizes (Hedges g), sample sizes, and methodology variables from validated full-text sources."
+capability: study_extraction + finding_extraction
 ---
+
 # Evidence Extraction Skill
 
-## 1. When to Use
+## When to Use
 Trigger on fetched and validated source texts to perform claim-level feature and statistical extraction.
 
-## 2. Process
-1. **Statistical Extraction**:
-   - Sample sizes (N_treatment, N_control).
-   - Means and standard deviations (M1, SD1, M2, SD2).
-   - Standardized effect size metric: compute Hedges' g, Cohen's d, or Odds Ratio.
-   - 95% Confidence Intervals [CI_lower, CI_upper] and p-values.
-2. **Methodology Extraction**:
-   - Design type: RCT, Quasi-Experimental (DID, PSM, RDD), Correlational.
-   - Outcome classification: Task Performance vs Conceptual Learning vs Delayed Retention.
-3. **Output Contract**: Emit evidence objects per schemas/evidence.schema.json（V1 顶层契约，修订 1.1）into evidence.jsonl; 图谱层投射为 EvidenceLink 时用 schemas/v2/evidence-link.schema.json（V2 契约）。
+## Inputs
+- `sources.jsonl`（全部 FETCH_VALID / 确认的 FETCH_PARTIAL）
+- `fetch/` 清正文；必要时经 Sciverse `/content` 读取定位段落
+
+## Process
+1. **Statistical Extraction**: sample sizes (N_treatment, N_control); means/SDs; standardized effect size (Hedges' g, Cohen's d, Odds Ratio via `scripts/effect_calculator.py`); 95% CIs and p-values when reported.
+2. **Methodology Extraction**: design type (RCT, quasi-experimental DID/PSM/RDD, correlational); outcome classification (task performance vs conceptual learning vs delayed retention).
+3. **Locate precisely**: keep `source_location` (page/section/offset) so every claim can be re-opened.
+
+## Output Contract
+Evidence Objects per `schemas/evidence.schema.json` (V1 top-level, revision 1.1) into `evidence.jsonl`; graph projections use `schemas/v2/evidence-link.schema.json` (V2).
+
+## Quality Gates
+- [ ] 每行通过 evidence schema，枚举合法。
+- [ ] 任务表现与学习效果记录分离。
+- [ ] 缺失统计量保持缺失（禁止由显著性反推效应量）。
+- [ ] 每条记录可定位回原文。
+
+## Anti-Patterns
+- 把 `relation_to_claim` 写到研究本体；把即测分数当保持/迁移；从摘要估算效应量。
+
+## Worked Example
+PNAS 2025 三臂 RCT → 三条 evidence：练习表现（task_performance, support）、独立考试（learning, contradict）、护栏组（learning, support），各带 N 与效应量。
+
+## References
+- `references/evidence-quality.md`、`references/effect_size_formulas.md`、`skill/task-briefs/extract.md`