eduevidence 5.2.0 → 6.2.0

package/docs/autoresearch-evolution-plan.md

@@ -0,0 +1,2903 @@
+# EduEvidence vNext — Autoresearch 自进化研究系统完整实施计划
+
+> 状态：Implementation Blueprint  
+> 范围：EduEvidence Skill + Research Engine + Benchmark + Agent MCP 编排 + Living Evidence  
+> 配套文档：[`docs/orchestration-role-model.md`](./orchestration-role-model.md)
+
+---
+
+# 0. 最终定义
+
+EduEvidence 下一阶段不增加一个新的“Autoresearch 用户功能页”，也不把现有研究流程改造成无限自主 Agent。
+
+保持当前三个用户入口：
+
+```text
+Evidence Review
+→ Decision & Pilot
+→ Evaluate & Update
+```
+
+保持 Canonical 9-step Protocol：
+
+```text
+Frame → Retrieve → Extract → Challenge → Audit → Adjudicate
+→ Applicability → Intervene → Evaluate
+```
+
+在它们下面增加一个 **Autonomous Research Meta-Layer**，形成三条不同性质的闭环：
+
+```text
+┌────────────────────────────────────────────────────────────┐
+│ Loop A — Evidence Autoresearch                             │
+│ “当前决策最值得继续寻找哪一条证据？”                        │
+│ Gap → Research Hypothesis → Experiment → Evidence Gain    │
+│ → GraphRevision → Decision Drift → Next Gap               │
+└────────────────────────────────────────────────────────────┘
+
+┌────────────────────────────────────────────────────────────┐
+│ Loop B — Decision-to-Outcome                              │
+│ “真实 Pilot 数据会不会改变当前决定？”                       │
+│ Decision → Pilot → Data → Analysis → GraphRevision        │
+│ → DecisionSnapshot                                        │
+└────────────────────────────────────────────────────────────┘
+
+┌────────────────────────────────────────────────────────────┐
+│ Loop C — Skill Autoresearch                               │
+│ “EduEvidence 这个研究系统本身还能不能做得更好？”             │
+│ Hypothesis → One Change → Eval → Keep/Revert → Repeat     │
+└────────────────────────────────────────────────────────────┘
+```
+
+三条循环对应三个问题：
+
+```text
+Evidence Loop：我们还缺什么知识？
+Decision Loop：新证据是否改变行动？
+Skill Loop：我们的研究方法本身还能不能更好？
+```
+
+最终产品定位升级为：
+
+> **Living Decision Research Engine**  
+> 不只把证据变成决定，还主动判断下一条最值得寻找的证据，并持续验证怎样让研究系统本身变得更好。
+
+---
+
+# 1. 核心思想来源
+
+## 1.1 Karpathy autoresearch：吸收“实验组织代码”，不复制训练任务
+
+`karpathy/autoresearch` 的关键设计不是“让 AI 一晚上跑很多次”，而是：
+
+```text
+固定 evaluator
+固定实验预算
+限制 mutation surface
+先跑 baseline
+一次只改一个核心变量
+真实运行
+机械测量
+keep / discard
+append-only 实验日志
+git 作为长期实验记忆
+```
+
+原版刻意把：
+
+```text
+prepare.py / evaluate_bpb
+```
+
+冻结，只允许 Agent 修改 `train.py`；每次训练使用固定 5 分钟时间预算，以单一 `val_bpb` 比较，并执行：
+
+```text
+commit
+→ run
+→ measure
+→ keep / reset
+→ log
+→ next experiment
+```
+
+其中 `program.md` 本质是“研究组织代码”：人类不是逐次编辑训练代码，而是定义 AI 研究员怎样组织自主实验。
+
+EduEvidence 直接吸收以下原则：
+
+1. **Evaluator 与研究对象分离。**
+2. **Mutation surface 有边界。**
+3. **Baseline 必跑。**
+4. **一次实验一个主假设。**
+5. **失败必须记录。**
+6. **结果弱而复杂度大，不保留。**
+7. **实验历史本身是未来推理输入。**
+8. **循环有预算、Plateau 和停止条件。**
+
+不照搬：
+
+```text
+单一 score 决定所有研究质量
+validated evidence keep/discard
+无限运行
+Agent 可改 evaluator
+```
+
+参考：
+- https://github.com/karpathy/autoresearch
+- https://github.com/karpathy/autoresearch/blob/master/program.md
+
+---
+
+## 1.2 通用 Autoresearch Skill：吸收“薄路由 + 可验证目标 + plateau”
+
+后续通用 autoresearch 实现进一步验证了几个适合 EduEvidence 的工程模式：
+
+```text
+薄 control-plane Skill
+子命令 / references 按需加载
+Success predicate
+Round-0 baseline/dry-run
+bounded iteration
+holdout / adversarial verify hop
+Plateau detection
+checkpoint + resumable state
+simplicity criterion
+```
+
+EduEvidence 不复制它的 14 个命令，但采用：
+
+```text
+Control Plane ≠ Detailed Procedures
+```
+
+这也与当前 Skill 的 progressive loading 方向一致。
+
+参考：
+- https://github.com/uditgoenka/autoresearch
+- https://github.com/uditgoenka/autoresearch/blob/master/guide/autoresearch-orchestrator.md
+
+---
+
+## 1.3 Living Systematic Review：Autoresearch 不能破坏系统综述纪律
+
+Cochrane 对 living systematic review 的核心定义是：
+
+> Review 持续更新，新证据出现后被识别并纳入，而不是每次重新从零生成一个结论。
+
+这与 EduEvidence 当前 `engine/living.py` 的：
+
+```text
+DecisionSnapshot subscription
+→ incremental evidence
+→ GraphRevision n+1
+→ drift
+```
+
+高度一致。
+
+Autoresearch 不替换 Living Evidence，而是让 Living Evidence 从：
+
+```text
+“监控预设 query terms”
+```
+
+升级为：
+
+```text
+“根据当前 Evidence Graph / KnowledgeGap / Decision Boundary
+动态决定下一轮最有价值的检索目标”
+```
+
+参考：
+- https://training.cochrane.org/handbook/current/chapter-22
+
+---
+
+## 1.4 Active Learning Screening：优化“先看什么”，不能决定“什么是真证据”
+
+ASReview 等 active-learning systematic review 工具证明：
+
+```text
+机器学习可以显著改善 record screening 的排序效率
+```
+
+EduEvidence 可以吸收为：
+
+```text
+Screening Priority
+```
+
+用于回答：
+
+> 下一篇最值得 Fetch / Screen 的候选 Source 是哪一篇？
+
+但严格禁止把 active-learning ranking 直接解释为：
+
+```text
+证据权重
+研究质量
+支持结论概率
+```
+
+Ranking 只优化 screening order。
+
+参考：
+- https://asreview.nl/project/read-the-docs/
+
+---
+
+## 1.5 Value of Information：把“下一步研究什么”连接到 Decision
+
+Value of Information 的核心不是“信息越多越好”，而是：
+
+> 新信息是否足以减少当前决策的不确定性，从而避免错误行动？
+
+完整 EVPI / EVPPI / EVSI 通常需要明确的 decision model、utility / loss 和概率分布。EduEvidence 当前没有资格默认生成这些正式统计量。
+
+因此 vNext 首先实现：
+
+```text
+Conceptual Decision Value of Information
+```
+
+简称：
+
+```text
+DVI — Decision Value of Information
+```
+
+它是**透明的研究优先级启发式**，不是概率、不是效应量、不是正式 EVSI。
+
+只有未来某个项目具备完整 decision-analytic model 时，才允许接入 formal VOI adapter。
+
+参考：
+- https://pubmed.ncbi.nlm.nih.gov/32113617/
+- https://pmc.ncbi.nlm.nih.gov/articles/PMC7613968/
+- https://pubmed.ncbi.nlm.nih.gov/25986471/
+
+---
+
+# 2. 三条“宪法级”原则
+
+保留现有：
+
+> **Optimize for decision integrity, not answer confidence.**
+
+新增：
+
+> **Optimize the research process, never the conclusion.**
+
+新增：
+
+> **Every iteration must improve the evidence state, improve the research system, or teach us why an attempted path failed.**
+
+这三条进入 `references/scientific-invariants.md`，并纳入 protected manifest。
+
+---
+
+# 3. 当前仓库的可复用地基
+
+本次不是重建。
+
+现有模块直接复用：
+
+| 当前能力 | 文件/目录 | vNext 用途 |
+|---|---|---|
+| Canonical 9-step | `docs/architecture.md` | 不变 |
+| Project/Run/Revision/DecisionSnapshot | `engine/` | 继续作为 canonical state |
+| Immutable Graph Revision | `engine/graph_store.py` | Evidence Autoresearch 的写入基础 |
+| KnowledgeGap derivation | `engine/gaps.py` | 加 DVI / lifecycle，不替换原推导 |
+| Study Design grounding | `engine/study_design.py` | Search saturation → Pilot 的硬 gate |
+| Pilot reinjection | `engine/pilot.py` / `engine/update.py` | Decision-to-Outcome Loop |
+| Living Evidence | `engine/living.py` | Evidence Autoresearch 的增量 ingestion 基础 |
+| Decision adjudication | `engine/tribunal.py` | 新 GraphRevision 后重算 |
+| EventBus | `engine/events.py` | Autoresearch 全过程事件流 |
+| Agent MCP adapter | `integrations/agent_mcp.py` | bounded worker dispatch |
+| `safe_spawn()` | `integrations/agent_mcp.py` | 唯一多代理派发入口继续保留 |
+| S/M/L Complexity | `scripts/complexity_gate.py` | ExecutionPlanner 输入 |
+| 30 题 Benchmark | `benchmarks/questions.jsonl` | DEV eval 起点 |
+| Gold annotations | `benchmarks/annotations/` | scientific eval 起点 |
+| Empirical benchmark | `benchmarks/empirical/` | Outer Loop 真实评估 |
+| Graph deterministic metrics | `benchmarks/evaluator/v2_graph_metrics.py` | L0 contract gates |
+| CI | `.github/workflows/ci.yml` | protected / autoresearch gate 扩展 |
+
+---
+
+# 4. 必须消除的架构混淆
+
+当前仓库同时存在：
+
+```text
+9 个 Protocol Stage
+8 个 skill/agents profiles
+多个 skill/sub-skills capabilities
+Agent MCP multi-model dispatch
+```
+
+vNext 唯一定义：
+
+```text
+Protocol Stage
+≠ Scientific Role
+≠ Capability
+≠ Worker/Subagent
+≠ Model/CLI
+```
+
+完整规则见：
+
+[`docs/orchestration-role-model.md`](./orchestration-role-model.md)
+
+核心一句：
+
+> **Parallelize independent evidence acquisition; serialize canonical state transitions.**
+
+以及：
+
+> **Subagent count is determined by task decomposability, not role count.**
+
+---
+
+# 5. 最终总体架构
+
+```text
+                          USER
+                            │
+                            ▼
+                    EduEvidence Skill
+                            │
+          Evidence Review / Decision & Pilot /
+                 Evaluate & Update
+                            │
+                            ▼
+                     Lead Researcher
+                            │
+               Workflow + Complexity Gate
+                            │
+                     Capability DAG
+                            │
+              ┌─────────────┴────────────┐
+              │                          │
+              ▼                          ▼
+       Platform Native            Agent MCP Enhanced
+         Local execution          bounded subagents
+              │                          │
+              └─────────────┬────────────┘
+                            ▼
+                    Staging Artifacts
+                            │
+            Schema / Provenance / Science Gates
+                            │
+                            ▼
+                    Single Writer Engine
+                            │
+              ┌─────────────┼─────────────┐
+              ▼             ▼             ▼
+        Evidence Graph  Decision State  Research Memory
+              │
+              ▼
+        GraphRevision N
+              │
+         ┌────┴───────────────┐
+         ▼                    ▼
+ Evidence Autoresearch   Decision-to-Outcome
+         │                    │
+         └─────────┬──────────┘
+                   ▼
+             GraphRevision N+1
+
+
+Repository / Skill Implementation
+              │
+              ▼
+        Skill Autoresearch
+ Hypothesis → Candidate → Protected Eval
+         → Keep/Revert → SkillRevision
+```
+
+---
+
+# 6. Canonical State 命名重新冻结
+
+避免“Revision”概念混乱。
+
+## 现实研究状态
+
+```text
+Project
+Run
+GraphRevision
+DecisionSnapshot
+ResearchIteration
+KnowledgeGap
+```
+
+### GraphRevision
+
+表示：
+
+> 我们对现实问题所拥有的 Evidence Graph 变了。
+
+例如：
+
+```text
+GraphRevision 7 → 8
+```
+
+---
+
+### DecisionSnapshot
+
+表示：
+
+> 基于某个 GraphRevision 的当前 decision-bounded adjudication。
+
+---
+
+### ResearchIteration
+
+表示：
+
+> Evidence Autoresearch 做过的一轮“下一条证据”实验。
+
+它可以：
+
+```text
+找到 valid evidence
+只找到重复 evidence
+没有找到 eligible evidence
+tool failure
+search saturation evidence
+```
+
+ResearchIteration 不等于 GraphRevision。
+
+没有新 validated evidence 时：
+
+```text
+ResearchIteration +1
+GraphRevision 不变
+```
+
+这是必须保留的信息。
+
+---
+
+## Skill 自身状态
+
+```text
+SkillRevision
+SkillExperiment
+EvalSnapshot
+```
+
+### SkillRevision
+
+表示：
+
+> EduEvidence 研究系统本身的一版实现。
+
+它与 GraphRevision 完全分离。
+
+---
+
+# 7. Loop A — Evidence Autoresearch
+
+## 7.1 目标
+
+现在的 Evidence Review 主要回答：
+
+```text
+当前证据支持什么？
+```
+
+Evidence Autoresearch 再回答：
+
+```text
+如果我们还允许花下一单位 research budget，
+最值得获取哪类新证据？
+```
+
+---
+
+## 7.2 基本循环
+
+```text
+READ STATE
+Project + GraphRevision N + DecisionSnapshot N
++ KnowledgeGaps + previous ResearchIterations
+        ↓
+RANK GAPS
+Decision Value of Information
+        ↓
+SELECT ONE
+one KnowledgeGap / one research objective
+        ↓
+HYPOTHESIZE
+one falsifiable research strategy
+        ↓
+PLAN
+TaskSpec + budget + termination
+        ↓
+EXECUTE
+Retrieve / Fetch / Validate / Extract
++ Challenge / Audit when triggered
+        ↓
+MEASURE
+Evidence Gain / Search Yield / Decision Relevance
+        ↓
+COMMIT
+valid evidence append-only → GraphRevision N+1
+or no-change iteration log
+        ↓
+RE-ADJUDICATE
+candidate DecisionSnapshot / drift
+        ↓
+LEARN
+append ResearchIteration + negative search memory
+        ↓
+NEXT
+recompute gap priority
+```
+
+---
+
+# 8. KnowledgeGap vNext
+
+当前 `engine/gaps.py` 的强项是：Gap 来自真实 graph coverage，而不是自由文本 future work。
+
+保留这个定义。
+
+新增 metadata：
+
+```json
+{
+  "gap_id": "GAP-...",
+  "status": "open",
+  "gap_type": "missing_transfer",
+  "derived_from_graph_revision": 8,
+  "decision_relevance": "high",
+  "uncertainty_source": "directness",
+  "minimum_useful_evidence": {
+    "population": "novice CS1",
+    "outcome": "independent_problem_solving",
+    "measurement_context": "without_ai"
+  },
+  "research_priority": {
+    "dvi_band": "high",
+    "rationale": []
+  },
+  "attempted_strategy_ids": [],
+  "negative_search_ids": [],
+  "resolution_status": "unresolved"
+}
+```
+
+---
+
+# 9. Gap Lifecycle
+
+新增 lifecycle：
+
+```text
+OPEN
+↓
+PRIORITIZED
+↓
+SEARCHING
+↓
+PARTIALLY_RESOLVED
+↓
+RESOLVED
+```
+
+另一条：
+
+```text
+OPEN
+↓
+SEARCHING
+↓
+SEARCH_SATURATED
+↓
+EMPIRICAL_EVIDENCE_NEEDED
+↓
+StudyDesign / Pilot
+```
+
+另一条：
+
+```text
+OPEN
+↓
+LOW_DECISION_VALUE
+```
+
+Low DVI 不代表“永远不研究”，只是当前 decision context 下不优先。
+
+---
+
+# 10. DVI — Decision Value of Information
+
+## 10.1 vNext 只实现 Conceptual DVI
+
+禁止输出：
+
+```text
+“这项研究有 83.7% 的信息价值”
+```
+
+也禁止把 DVI 称为正式 EVPI / EVSI。
+
+DVI 是透明优先级模型。
+
+---
+
+## 10.2 排序维度
+
+每个 Gap 计算以下 ordinal dimensions：
+
+```text
+Decision Sensitivity
+当前 decision 对这个未知量有多敏感？
+
+Current Uncertainty
+当前 evidence 对它有多不确定？
+
+Directness Deficit
+缺的是直接 outcome 证据还是仅补充性信息？
+
+Applicability Relevance
+是否直接关系目标人群 / 场景？
+
+Expected Evidence Availability
+二手证据是否可能存在？
+
+Research Cost
+预计检索/筛选/新研究成本？
+
+Risk / Irreversibility
+如果现在做错决定，后果是否高？
+```
+
+---
+
+## 10.3 结果只分 band
+
+```text
+HIGH
+MEDIUM
+LOW
+```
+
+同时返回 rationale：
+
+```json
+{
+  "gap_id": "GAP-...",
+  "dvi_band": "HIGH",
+  "drivers": [
+    "当前 ADOPT/PILOT 边界主要被 independent transfer 缺失限制",
+    "目标人群完全匹配",
+    "已有候选文献线索"
+  ],
+  "cost_band": "LOW",
+  "next_research_mode": "secondary_evidence_search"
+}
+```
+
+排序必须可解释。
+
+---
+
+# 11. 一轮只能有一个主 Research Hypothesis
+
+示例：
+
+```json
+{
+  "research_iteration_id": "RIT-0017",
+  "gap_id": "GAP-transfer-novice",
+  "hypothesis": "针对 novice CS1 + without-AI transfer test 的定向检索，将发现至少一条比当前证据更直接的 eligible Study",
+  "strategy_id": "STRAT-targeted-transfer-v1",
+  "expected_gain": "directness",
+  "budget": {
+    "max_queries": 6,
+    "max_candidates": 30,
+    "max_fulltext_fetches": 12
+  }
+}
+```
+
+禁止一轮同时：
+
+```text
+换检索算法
+换 OutcomeClassifier
+改 Gap ranking
+加新数据库
+改 Method Audit prompt
+```
+
+否则无法归因。
+
+---
+
+# 12. Research Experiment 类型
+
+首版只允许明确枚举：
+
+```text
+TARGETED_RETRIEVAL
+COUNTER_EVIDENCE_RETRIEVAL
+APPLICABILITY_RETRIEVAL
+TEMPORAL_REFRESH
+CITATION_CHAINING
+SCREENING_PRIORITY
+SOURCE_RECOVERY
+```
+
+未来再增加。
+
+它们都是：
+
+> 获取证据的研究策略实验。
+
+不是：
+
+> 自动改变科学规则。
+
+---
+
+# 13. Evidence 绝对 Append-only
+
+这是从 autoresearch 移植时最重要的差异。
+
+代码候选可以：
+
+```text
+keep / revert
+```
+
+Validated Evidence 不可以：
+
+```text
+支持当前结论 → keep
+反对当前结论 → discard
+```
+
+只要符合：
+
+```text
+真实来源
+符合纳排标准
+provenance 可验证
+schema 合法
+```
+
+就必须保留，不论：
+
+```text
+support
+contradict
+neutral
+null
+```
+
+科学不变量：
+
+> **Never optimize the evidence set toward a preferred conclusion.**
+
+---
+
+# 14. Negative Search Result 也是 Artifact
+
+新增：
+
+```text
+NegativeSearchRecord
+```
+
+记录：
+
+```json
+{
+  "negative_search_id": "NSR-...",
+  "research_iteration_id": "RIT-...",
+  "gap_id": "GAP-...",
+  "queries": [],
+  "providers": [],
+  "candidate_count": 23,
+  "fetched_count": 8,
+  "eligible_count": 0,
+  "exclusion_reasons": {},
+  "searched_at": "...",
+  "scope": {},
+  "conclusion": "no_eligible_evidence_found_within_search_scope"
+}
+```
+
+禁止写：
+
+```text
+“No evidence exists.”
+```
+
+只能说：
+
+```text
+“在本轮明确 search scope 内未找到 eligible evidence。”
+```
+
+---
+
+# 15. Search Saturation Gate
+
+不能无限搜论文。
+
+每个 Gap 追踪：
+
+```text
+new independent studies per iteration
+new eligible findings per iteration
+new direct-outcome findings
+new population coverage
+new contradiction yield
+new DOI/source uniqueness
+Decision Boundary delta
+query/provider novelty
+```
+
+当连续多轮出现：
+
+```text
+unique eligible evidence ≈ 0
+直接证据增量 = 0
+重复来源持续升高
+Decision Boundary 无变化
+新的 search strategies 已基本耗尽
+```
+
+标记：
+
+```text
+SEARCH_SATURATED
+```
+
+首版不用一个武断的单阈值。
+
+使用：
+
+```text
+2–3 consecutive low-yield iterations
++ strategy diversity exhausted
++ no high-value unresolved retrievable path
+```
+
+作为 deterministic + policy 判定。
+
+---
+
+# 16. 自动从 Literature 转向 Pilot 的条件
+
+必须同时满足：
+
+```text
+Gap DVI = HIGH
+Gap 对 Decision material
+Secondary Search = saturated
+Gap 仍 unresolved
+新实证研究在伦理 / 可行性上允许
+```
+
+然后：
+
+```text
+KnowledgeGap
+→ EMPIRICAL_EVIDENCE_NEEDED
+→ StudyDesign grounding gate
+→ preregistration-ready Pilot
+```
+
+这直接复用当前：
+
+```text
+No Study Design Without Grounded Gap
+```
+
+---
+
+# 17. Evidence Autoresearch 的停止条件
+
+任何一项满足即可停：
+
+```text
+目标 Gap RESOLVED
+Decision 对该 Gap 不再敏感
+所有 HIGH-DVI gaps 已处理
+research budget exhausted
+search saturation
+需要新的 empirical study
+用户要求停止
+工具不可用且无有效降级路径
+```
+
+不允许“NEVER STOP”。
+
+研究系统必须 bounded。
+
+---
+
+# 18. Decision 的自动化边界
+
+Evidence Autoresearch 可以自动：
+
+```text
+生成新的 GraphRevision
+运行 deterministic tribunal
+生成 candidate DecisionSnapshot / DriftReport
+```
+
+但不得自动：
+
+```text
+执行真实世界干预
+改变学校政策
+开始人类受试 Pilot
+对高影响 Decision 自动 promote 为 accepted operational decision
+```
+
+新增：
+
+```text
+DecisionSnapshot.status
+= candidate | reviewed | accepted
+```
+
+若迁移成本过高，首版可先在 `extensions.review_status` 实现，再升级 schema。
+
+Living refresh 生成：
+
+```text
+candidate
+```
+
+用户 / human gate 接受后：
+
+```text
+accepted
+```
+
+---
+
+# 19. Loop B — Decision-to-Outcome 保持现有语义
+
+不重写 `engine/pilot.py`。
+
+保持：
+
+```text
+Decision
+→ grounded StudyDesign
+→ PilotRun
+→ privacy-safe Data
+→ validated AnalysisRun
+→ local Finding
+→ GraphRevision
+→ DecisionSnapshot
+```
+
+Autoresearch 只负责在前面回答：
+
+> 为什么现在应该从 literature 转入这个 Pilot？
+
+---
+
+# 20. Loop C — Skill Autoresearch
+
+## 20.1 目标
+
+自动实验：
+
+> 怎样让 EduEvidence 的真实研究质量更高，同时不破坏科学纪律、鲁棒性和成本？
+
+---
+
+## 20.2 基本循环
+
+```text
+BASELINE
+↓
+Read previous experiments
+↓
+Propose ONE hypothesis
+↓
+Create candidate branch/worktree
+↓
+Modify allowed surface
+↓
+Fast deterministic gates
+↓
+DEV benchmark
+↓
+If promising → repeated empirical eval
+↓
+HOLDOUT / adversarial verify
+↓
+Promotion policy
+├─ KEEP → candidate best branch
+├─ REJECT → discard candidate code
+├─ RETEST → more repeats
+└─ HUMAN_REVIEW → ambiguous/high-impact
+↓
+Log everything
+↓
+Next experiment
+```
+
+---
+
+# 21. Protected Scientific Core
+
+新增：
+
+```text
+autoevolve/protected.manifest.yaml
+```
+
+首版保护：
+
+```yaml
+protected:
+  - benchmarks/annotations/**
+  - benchmarks/holdout/**
+  - benchmarks/evaluator/**
+  - schemas/**
+  - references/scientific-invariants.md
+  - scripts/pre_verdict_gate.py
+  - scripts/compute_confidence.py
+  - engine/graph_store.py::revision_integrity
+  - engine/study_design.py::grounding_gate
+  - tests/scientific_invariants/**
+```
+
+Runner 在实验前后计算 hash。
+
+任何 protected mutation：
+
+```text
+EXPERIMENT_INVALID
+```
+
+不进入评分。
+
+---
+
+# 22. Mutation Surface
+
+## Tier A — Safe Mutable
+
+适合自动实验：
+
+```text
+retrieval query templates
+search ordering
+Challenge prompts
+context packing
+role instructions
+workflow wording
+few-shot examples
+report phrasing
+routing heuristics that do not alter scientific invariants
+model routing / cost strategy
+cache strategy
+```
+
+---
+
+## Tier B — Controlled Mutable
+
+需要完整 regression：
+
+```text
+OutcomeClassifier implementation
+ranking weights
+Complexity Gate thresholds
+Gap priority heuristics
+Graph linking heuristics
+retrieval dedupe heuristics
+orchestration policy
+```
+
+---
+
+## Tier C — Protected
+
+不可自动修改：
+
+```text
+Gold / Holdout
+Evaluator definitions
+Schema ground truth
+No Direct Learning Evidence → No ADOPT
+No False Precision
+Fail-closed causal analysis
+Synthetic/real separation
+StudyDesign grounding
+privacy / ethics constraints
+canonical revision integrity
+```
+
+---
+
+# 23. Outer Loop 不是单分数优化
+
+Karpathy 原版可使用：
+
+```text
+val_bpb ↓
+```
+
+EduEvidence 必须采用：
+
+```text
+Constraint-first Pareto Evaluation
+```
+
+---
+
+# 24. 评估层级
+
+## L0 — Scientific / Contract Hard Gates
+
+任意失败：
+
+```text
+REJECT
+```
+
+包括：
+
+```text
+Schema validity
+Graph traceability
+Study identity correctness
+Independent study counting
+No false precision
+No unsupported ADOPT
+Synthetic-data separation
+Grounded Gap requirement
+Protected file integrity
+PII / privacy guards
+```
+
+当前 `v2_graph_metrics.py` 继续作为其中一部分。
+
+它明确只能测 software / contract conformance，不能声称 scientific truth；该边界保持。
+
+---
+
+## L1 — Scientific Correctness
+
+```text
+Outcome Separation
+Citation Support
+Contradiction precision/recall
+Decision Calibration
+Scope Calibration
+Methodology issue detection
+Gap correctness
+```
+
+使用：
+
+```text
+gold annotations
+human/judge annotations where appropriate
+```
+
+---
+
+## L2 — Research Quality
+
+新增：
+
+```text
+Direct Evidence Gain
+Unique Eligible Evidence Yield
+Counter-evidence Yield
+Applicability Coverage
+Gap Resolution Rate
+Research Saturation Efficiency
+Decision-relevant Evidence Gain
+```
+
+---
+
+## L3 — Robustness
+
+```text
+S/M/L
+multiple domains
+multiple model families
+repeated runs
+retrieval provider variation
+missing-data scenarios
+adversarial cases
+```
+
+---
+
+## L4 — Efficiency
+
+```text
+token
+cost
+latency
+search calls
+fetch calls
+subagent count
+parallel speedup
+```
+
+---
+
+## L5 — Simplicity
+
+```text
+LOC delta
+SKILL.md token load
+number of rules
+number of branches
+new dependencies
+maintenance surface
+```
+
+同等质量：
+
+```text
+更简单 = 更优
+```
+
+---
+
+# 25. Promotion Policy
+
+先看 hard gates。
+
+```text
+if L0 fails:
+    REJECT
+```
+
+然后：
+
+```text
+if core scientific metric material regression:
+    REJECT
+```
+
+否则比较 Pareto：
+
+```text
+quality
+robustness
+efficiency
+simplicity
+```
+
+状态：
+
+```text
+KEEP
+REJECT
+RETEST
+HUMAN_REVIEW
+CRASH
+INVALID
+```
+
+---
+
+# 26. Noise Floor 与重复运行
+
+LLM 输出不是 deterministic benchmark。
+
+保持当前 benchmark 的：
+
+```text
+repeats >= 3
+prefer 5
+mean + variance / CI
+```
+
+并进一步：
+
+```text
+小于 noise floor 的提升不得自动 KEEP
+```
+
+候选状态：
+
+```text
+candidate delta < empirical noise
+→ RETEST
+```
+
+不要让 +0.2% 的偶然波动累计成“自我进化”。
+
+---
+
+# 27. Benchmark 三分区
+
+当前 30 题不能让 Agent 反复看到全部内容。
+
+重组：
+
+```text
+benchmarks/
+├── dev/
+├── holdout/
+├── adversarial/
+├── temporal/
+└── evaluator/
+```
+
+---
+
+## DEV
+
+Agent 可以读取。
+
+用途：
+
+```text
+fast iteration
+```
+
+---
+
+## HOLDOUT
+
+Candidate 不能读取题目与 gold。
+
+只允许 protected evaluator 调用。
+
+用途：
+
+```text
+promotion gate
+```
+
+---
+
+## ADVERSARIAL
+
+至少覆盖：
+
+```text
+fake DOI
+fabricated citation
+search snippet pretending to be evidence
+missing CI
+same Study split into multiple Findings
+task performance treated as learning
+population mismatch
+subgroup overclaim
+contradictory evidence
+retracted / corrected source
+synthetic evidence leakage
+prompt injection in source content
+PII in empirical dataset
+singular DID / invalid causal design
+```
+
+---
+
+## TEMPORAL
+
+用于：
+
+```text
+recent evidence refresh
+new publications
+model/provider drift
+```
+
+不能 hard-code 成永久 gold；保留 evaluation timestamp。
+
+---
+
+# 28. Benchmark Hacking Guard
+
+任何实验不得：
+
+```text
+修改 evaluator
+读取 holdout gold
+减少 benchmark workload
+跳过失败题
+把 unknown 当 pass
+针对 question ID 写特例
+通过少输出降低 unsupported rate
+```
+
+防止“少说少错”：
+
+```text
+precision + recall 必须成对报告
+```
+
+例如：
+
+```text
+Contradiction precision
+Contradiction recall
+```
+
+同时看。
+
+---
+
+# 29. Experiment Memory
+
+新增：
+
+```text
+autoevolve/
+├── program.md
+├── config.yaml
+├── protected.manifest.yaml
+├── results.tsv
+├── experiments.jsonl
+├── best.json
+├── ideas.md
+├── findings.md
+└── runs/
+```
+
+---
+
+# 30. `results.tsv`
+
+```text
+experiment_id
+parent_revision
+candidate_commit
+scope
+hypothesis
+eval_suite_hash
+model_manifest
+repeats
+hard_gates
+science_score
+research_score
+robustness
+cost
+latency
+complexity_delta
+status
+description
+```
+
+失败同样写入。
+
+`results.tsv` 是 append-only。
+
+---
+
+# 31. `experiments.jsonl`
+
+每轮保留完整机器可读结构：
+
+```json
+{
+  "experiment_id": "EXP-0017",
+  "hypothesis": "显式拆分 direct evidence 与 transfer query 能提高 contradiction/directness recall",
+  "why": "E12/E15 显示当前 retrieve 在 outcome axis 上混合",
+  "change": [],
+  "baseline": {},
+  "candidate": {},
+  "result": "KEEP",
+  "interpretation": "...",
+  "what_this_rules_out": [],
+  "next_ideas": []
+}
+```
+
+Research Memory 不只保存 score。
+
+---
+
+# 32. Git Workflow
+
+Autoresearch 永远不直接在 `main` 上实验。
+
+```text
+main
+└── autoresearch/<run-tag>
+    ├── experiment commit 1
+    ├── experiment commit 2
+    └── best candidate
+```
+
+更稳妥实现：
+
+```text
+git worktree
+```
+
+每个 candidate 有独立工作树。
+
+---
+
+# 33. 自动 Promotion 边界
+
+自动允许：
+
+```text
+在 autoresearch branch 内 keep / revert
+更新 best candidate pointer
+生成 PR 内容
+```
+
+自动禁止：
+
+```text
+merge main
+release
+修改 protected core
+删除 benchmark history
+force push
+真实部署
+```
+
+最终 merge 人工确认。
+
+---
+
+# 34. Plateau / Ceiling
+
+Skill Autoresearch 不无限运行。
+
+默认：
+
+```text
+max_experiments_per_session = 25
+```
+
+每日 unattended 可配置：
+
+```text
+max_experiments = 50
+max_cost
+max_wall_time
+```
+
+Plateau：
+
+```text
+连续 5 个 valid experiments
+没有超过 noise floor 的 Pareto improvement
+```
+
+停止并输出：
+
+```text
+plateau report
+best revision
+near misses
+next hypotheses
+```
+
+---
+
+# 35. 每日自动进化模式
+
+未来 CLI：
+
+```bash
+eduevidence evolve run --budget daily
+```
+
+配置：
+
+```yaml
+schedule_profile: daily
+max_experiments: 20
+max_cost_usd: 5
+max_wall_minutes: 180
+mutation_tiers:
+  - safe
+allow_controlled: false
+promotion: branch_only
+```
+
+每次运行结束：
+
+```text
+Daily Evolution Report
+├── baseline
+├── experiments attempted
+├── kept
+├── rejected
+├── crashes
+├── best candidate
+├── metric deltas
+├── cost
+├── protected integrity
+└── next ideas
+```
+
+初版不让 daily mode 自动改 Tier B。
+
+---
+
+# 36. 多角色 / 子代理编排
+
+完整权威定义：
+
+[`docs/orchestration-role-model.md`](./orchestration-role-model.md)
+
+核心改变：
+
+```text
+8 role profiles
+≠
+8 permanent agents
+```
+
+---
+
+# 37. 新增 Lead Researcher / ExecutionPlanner
+
+逻辑组件：
+
+```text
+engine/orchestration/planner.py
+```
+
+职责：
+
+```text
+Workflow routing
+Complexity classification
+Capability DAG
+local vs delegate decision
+parallel groups
+TaskSpec creation
+budget allocation
+independence requirements
+fallback
+```
+
+它不直接调用外部 daemon。
+
+执行仍走：
+
+```text
+integrations/agent_mcp.py::safe_spawn()
+```
+
+---
+
+# 38. TaskSpec
+
+新增 schema：
+
+```text
+schemas/vNext/task-spec.schema.json
+```
+
+必须包括：
+
+```text
+task_id
+run_id
+base_revision
+role_profile
+objective
+reason_for_delegation
+input_artifacts
+allowed_capabilities
+forbidden_actions
+scope
+budget
+output_contract
+termination
+```
+
+没有 TaskSpec 不 spawn。
+
+---
+
+# 39. Single Writer
+
+任何 Worker 只能输出：
+
+```text
+Staging Artifact
+```
+
+只有主引擎可以：
+
+```text
+Graph commit
+DecisionSnapshot write
+KnowledgeGap persistent update
+StudyDesign write
+```
+
+并行只发生在证据获取 / 分析。
+
+Canonical State transition 串行。
+
+---
+
+# 40. 默认 S/M/L 拓扑
+
+## S
+
+```text
+Lead only
+subagents = 0
+```
+
+---
+
+## M
+
+典型：
+
+```text
+Lead
+├─ direct/support retrieval worker
+├─ counter/risk retrieval worker
+├─ optional Method Reviewer
+└─ Judge / Lead adjudication
+```
+
+典型 active workers：
+
+```text
+2–4
+```
+
+---
+
+## L
+
+```text
+Lead
+│
+├─ Retrieval Worker — direct causal
+├─ Retrieval Worker — transfer/retention
+├─ Retrieval Worker — harms/null/counter
+├─ Retrieval Worker — applicability/freshness
+│
+├─ deterministic merge/dedupe
+├─ optional Extraction batches
+│
+├─ independent Skeptic
+├─ Method Reviewer
+│
+├─ Single Writer Graph commit
+├─ Evidence Judge
+└─ high-impact Independent Final Review
+```
+
+默认：
+
+```text
+parallel active workers <= 6
+```
+
+禁止 recursive swarm。
+
+---
+
+# 41. Worker 按 Evidence Axis 拆，不按 Provider 拆
+
+优先：
+
+```text
+直接效果
+transfer / retention
+negative / null / risks
+population / subgroup
+current/fresh evidence
+```
+
+而不是：
+
+```text
+Google Worker
+Crossref Worker
+Semantic Scholar Worker
+```
+
+Provider 是工具，不是 epistemic objective。
+
+---
+
+# 42. Independence 的三种含义
+
+```text
+Search independence
+Context independence
+Model-family independence
+```
+
+只在需要时使用。
+
+Skeptic / final reviewer 才优先强调：
+
+```text
+context + model-family independence
+```
+
+Retriever workers 不必浪费成本全部使用不同旗舰模型。
+
+---
+
+# 43. Multi-agent Autoresearch
+
+Outer Loop 可以自动实验：
+
+```text
+1 vs 2 vs 4 retrieval workers
+provider split vs evidence-axis split
+Skeptic always vs triggered
+Method Review batch size
+TaskSpec detail level
+worker context bundle size
+S/M/L spawn thresholds
+fast/strong routing
+```
+
+优化目标：
+
+```text
+Marginal Evidence Gain / Added Cost
+```
+
+不是：
+
+```text
+Agent 数量越多越高级
+```
+
+---
+
+# 44. Orchestration Metrics
+
+新增：
+
+```text
+Duplicate Search Rate
+Unique Eligible Evidence / Worker
+Coverage Gain / Worker
+Counter Evidence Yield
+Artifact Acceptance Rate
+Worker Failure Recovery Rate
+Coordination Token Cost
+Parallel Speedup
+Canonical State Conflict
+Unvalidated Text Used by Judge
+```
+
+硬目标：
+
+```text
+Canonical State Conflict = 0
+Unvalidated Text Used by Judge = 0
+```
+
+---
+
+# 45. EventBus 新事件
+
+## Evidence Autoresearch
+
+```text
+autoresearch.evidence.started
+autoresearch.gap.ranked
+autoresearch.hypothesis.created
+autoresearch.iteration.started
+autoresearch.search.completed
+autoresearch.iteration.no_gain
+autoresearch.iteration.evidence_gain
+autoresearch.saturation.detected
+autoresearch.empirical_needed
+autoresearch.evidence.completed
+```
+
+## Skill Autoresearch
+
+```text
+autoevolve.session.started
+autoevolve.experiment.created
+autoevolve.candidate.built
+autoevolve.eval.completed
+autoevolve.candidate.kept
+autoevolve.candidate.rejected
+autoevolve.candidate.retest
+autoevolve.plateau
+autoevolve.session.completed
+```
+
+## Orchestration
+
+```text
+orchestrator.plan.created
+orchestrator.task.dispatched
+orchestrator.task.completed
+orchestrator.task.failed
+orchestrator.artifact.accepted
+orchestrator.artifact.rejected
+```
+
+---
+
+# 46. CLI 设计
+
+不破坏现有命令。
+
+新增两个命令域：
+
+```text
+research auto
+
+evolve
+```
+
+---
+
+## Evidence Autoresearch
+
+```bash
+eduevidence research auto start --project PRJ-...
+eduevidence research auto step  --project PRJ-...
+eduevidence research auto status --project PRJ-...
+eduevidence research auto report --project PRJ-...
+eduevidence research auto stop --project PRJ-...
+```
+
+`step` 永远只执行一轮，便于测试。
+
+`start` 才执行 bounded loop。
+
+---
+
+## Skill Autoresearch
+
+```bash
+eduevidence evolve init
+eduevidence evolve baseline
+eduevidence evolve run --max-experiments 20
+eduevidence evolve status
+eduevidence evolve report
+eduevidence evolve best
+eduevidence evolve prepare-pr
+```
+
+不提供：
+
+```text
+auto-merge-main
+```
+
+---
+
+# 47. 新目录
+
+```text
+eduevidence/
+├── autoevolve/
+│   ├── program.md
+│   ├── config.yaml
+│   ├── protected.manifest.yaml
+│   ├── results.tsv
+│   ├── experiments.jsonl
+│   ├── best.json
+│   └── runs/
+│
+├── engine/
+│   ├── autoresearch/
+│   │   ├── controller.py
+│   │   ├── gap_priority.py
+│   │   ├── hypothesis.py
+│   │   ├── iteration.py
+│   │   ├── saturation.py
+│   │   ├── research_memory.py
+│   │   └── contracts.py
+│   │
+│   ├── orchestration/
+│   │   ├── planner.py
+│   │   ├── role_registry.py
+│   │   ├── task_spec.py
+│   │   ├── worker_result.py
+│   │   └── merge.py
+│   │
+│   └── autoevolve/
+│       ├── controller.py
+│       ├── mutation_scope.py
+│       ├── evaluator.py
+│       ├── promotion.py
+│       ├── experiment_log.py
+│       └── git_workspace.py
+│
+├── skill/
+│   ├── roles/
+│   │   └── registry.yaml
+│   └── agents/                 # compatibility during migration
+│
+├── benchmarks/
+│   ├── dev/
+│   ├── holdout/
+│   ├── adversarial/
+│   ├── temporal/
+│   └── evaluator/
+│
+└── schemas/vNext/
+```
+
+---
+
+# 48. 新 Schema
+
+首版：
+
+```text
+research-iteration.schema.json
+research-strategy.schema.json
+negative-search-record.schema.json
+gap-priority.schema.json
+task-spec.schema.json
+worker-result.schema.json
+execution-plan.schema.json
+skill-experiment.schema.json
+eval-snapshot.schema.json
+autoevolve-session.schema.json
+```
+
+不立即重写原 v2/v3/v4 schema。
+
+---
+
+# 49. ResearchIteration Schema 核心字段
+
+```text
+iteration_id
+project_id
+base_graph_revision
+gap_id
+hypothesis
+strategy
+budget
+execution_plan_id
+search_attempts
+candidate_sources
+validated_evidence_ids
+negative_search_ids
+evidence_gain
+new_graph_revision|null
+decision_snapshot_id|null
+status
+started_at
+completed_at
+```
+
+状态：
+
+```text
+completed_gain
+completed_no_gain
+search_saturated
+empirical_needed
+budget_exhausted
+tool_failure
+invalid
+```
+
+---
+
+# 50. SkillExperiment Schema 核心字段
+
+```text
+experiment_id
+session_id
+parent_skill_revision
+hypothesis
+mutation_scope
+changed_files
+candidate_commit
+baseline_eval_id
+candidate_eval_id
+protected_hash_before
+protected_hash_after
+status
+promotion_reason
+complexity_delta
+```
+
+---
+
+# 51. Dashboard / Web Studio
+
+不是 P0。
+
+在 Engine 稳定后新增两个面板：
+
+## Research Loop
+
+```text
+Current Decision
+Open Gaps ranked by DVI
+Current Research Hypothesis
+Iterations
+Evidence Gain
+Search Saturation
+Decision Drift
+```
+
+## Skill Evolution
+
+```text
+Baseline vs Best
+Experiment Timeline
+Keep / Reject
+Metric Radar
+Cost
+Plateau
+Protected Core status
+```
+
+不要做“8 个 Agent 头像同时工作”。
+
+展示：
+
+```text
+Capability DAG + active TaskSpec
+```
+
+---
+
+# 52. SKILL.md 改造原则
+
+不要把整套 Autoresearch 写进顶层 `SKILL.md`。
+
+SKILL.md 只加入：
+
+```text
+Autonomous Research Meta-Layer
+```
+
+约 20–35 行，说明：
+
+```text
+何时启用 Evidence Autoresearch
+何时启用 Skill Autoresearch（仅维护者/开发模式）
+科学不变量
+references 路由
+```
+
+详细内容进入：
+
+```text
+references/autoresearch.md
+references/orchestration.md
+references/evaluation-policy.md
+```
+
+保持 SKILL.md control-plane。
+
+---
+
+# 53. Skill 自进化绝不能改用户研究状态
+
+Skill Autoresearch 使用 fixture / benchmark project。
+
+不得对真实用户 Project：
+
+```text
+自动修改 Graph
+生成新的 accepted Decision
+删除证据
+重写研究历史
+```
+
+Outer Loop 与用户数据物理隔离。
+
+---
+
+# 54. Evidence Autoresearch 绝不能改 Skill
+
+相反：
+
+Evidence Autoresearch 只可以影响：
+
+```text
+Project research state
+ResearchIteration log
+GraphRevision
+candidate DecisionSnapshot
+```
+
+不能修改 repo：
+
+```text
+SKILL.md
+retrieval prompts
+engine code
+evaluator
+```
+
+两条 loop 权限完全分离。
+
+---
+
+# 55. Agent MCP 权限
+
+继续遵守：
+
+```text
+Detect
+→ Recommend
+→ explicit user approval
+→ safe_spawn
+```
+
+Autoresearch 不绕过用户已经建立的模型 / CLI authorization。
+
+在 unattended Skill Autoresearch session 中，只能使用：
+
+```text
+session 开始前已批准的 model/CLI allowlist
+```
+
+不能中途自行授权新模型。
+
+---
+
+# 56. 安全与研究诚信
+
+硬规则：
+
+```text
+No evaluator mutation
+No holdout leakage
+No evidence deletion because unfavorable
+No fake precision
+No fake citations
+No autonomous human-subject study launch
+No PII upload
+No automatic main merge
+No recursive unbounded agent spawning
+No evidence-set optimization toward desired verdict
+```
+
+---
+
+# 57. Tests — Evidence Autoresearch
+
+新增：
+
+```text
+tests/test_autoresearch_gap_priority.py
+tests/test_autoresearch_iteration.py
+tests/test_autoresearch_negative_search.py
+tests/test_autoresearch_saturation.py
+tests/test_autoresearch_transition_to_pilot.py
+tests/test_autoresearch_append_only.py
+tests/test_autoresearch_decision_drift.py
+```
+
+必须覆盖：
+
+```text
+negative evidence 不被丢弃
+no-result 不被写成 no evidence exists
+no-gain iteration 不产生 GraphRevision
+重复 evidence 不产生 revision
+high-DVI saturated gap → empirical_needed
+low-DVI gap 不自动生成 Pilot
+```
+
+---
+
+# 58. Tests — Orchestration
+
+新增：
+
+```text
+tests/test_execution_planner.py
+tests/test_task_spec.py
+tests/test_worker_merge.py
+tests/test_single_writer.py
+tests/test_role_registry.py
+tests/test_orchestration_budget.py
+```
+
+硬测：
+
+```text
+S task = 0 workers by default
+M can selectively spawn
+L <= configured worker cap
+worker cannot graph_commit
+Judge only consumes validated artifacts
+no direct spawn bypasses safe_spawn
+```
+
+---
+
+# 59. Tests — Skill Autoresearch
+
+新增：
+
+```text
+tests/test_autoevolve_protected_manifest.py
+tests/test_autoevolve_experiment_log.py
+tests/test_autoevolve_promotion.py
+tests/test_autoevolve_noise_floor.py
+tests/test_autoevolve_plateau.py
+tests/test_autoevolve_holdout_isolation.py
+tests/test_autoevolve_git_workspace.py
+```
+
+---
+
+# 60. CI 扩展
+
+现有三 job 继续。
+
+新增：
+
+```text
+scientific-invariants
+protected-manifest
+orchestration-contract
+```
+
+PR 来自 `autoresearch/*` 时额外运行：
+
+```text
+DEV regression
+holdout promotion gate
+adversarial gate
+complexity report
+```
+
+Empirical model eval 成本较高：
+
+```text
+不放在每个普通 commit
+```
+
+只在：
+
+```text
+manual workflow_dispatch
+nightly / scheduled environment
+candidate promotion
+```
+
+运行。
+
+---
+
+# 61. Phase 0 — Architecture Freeze
+
+交付：
+
+```text
+docs/autoresearch-evolution-plan.md
+docs/orchestration-role-model.md
+references/scientific-invariants.md
+```
+
+冻结：
+
+```text
+3 loops
+5 orchestration concepts
+Single Writer
+protected evaluator
+append-only evidence
+```
+
+验收：
+
+所有核心文档不再把 8 role 写成 8 常驻 Agents。
+
+---
+
+# 62. Phase 1 — Orchestration Clarity
+
+实现：
+
+```text
+skill/roles/registry.yaml
+ExecutionPlanner
+TaskSpec
+WorkerResult
+ExecutionPlan artifact
+```
+
+先不做 Autoresearch。
+
+原因：
+
+> 不能在不清晰的 agent topology 上建立自主循环。
+
+验收：
+
+```text
+S/M/L deterministic plan tests
+Agent MCP approval tests 全通过
+safe_spawn 仍是唯一 spawn path
+```
+
+---
+
+# 63. Phase 2 — Evidence Research Memory
+
+实现：
+
+```text
+ResearchIteration
+NegativeSearchRecord
+ResearchStrategy
+append-only iteration history
+```
+
+此阶段仍由人工选择 Gap。
+
+CLI：
+
+```bash
+eduevidence research auto step --gap GAP-...
+```
+
+验收：
+
+```text
+一轮可完整复现
+no-gain 正确记录
+Graph 只在 valid evidence 时 revision
+```
+
+---
+
+# 64. Phase 3 — Gap Priority + DVI
+
+扩展 `engine/gaps.py`，不替换现有 coverage derivation。
+
+新增：
+
+```text
+DVI band
+minimum useful evidence
+researchability
+research cost band
+```
+
+验收 fixture：
+
+CS1 案例中：
+
+```text
+missing independent transfer
+```
+
+应优先于：
+
+```text
+低 decision relevance 的满意度缺口
+```
+
+同时必须给出可读 rationale。
+
+---
+
+# 65. Phase 4 — Bounded Evidence Autoresearch
+
+实现 controller：
+
+```text
+rank
+→ select
+→ hypothesize
+→ dispatch
+→ validate
+→ measure
+→ log
+→ next
+```
+
+首版：
+
+```text
+max_iterations = 5
+```
+
+默认不是无限。
+
+验收：
+
+```text
+可以跑完整 3–5 iteration
+worker failure 可恢复
+不重复同一个失败策略
+```
+
+---
+
+# 66. Phase 5 — Search Saturation → Pilot Bridge
+
+实现：
+
+```text
+saturation.py
+```
+
+然后连接：
+
+```text
+SEARCH_SATURATED
++ HIGH DVI
+→ EMPIRICAL_EVIDENCE_NEEDED
+→ existing StudyDesign gate
+```
+
+验收：
+
+```text
+不能直接凭“资料少”生成 Pilot
+必须有 Gap + attempts + saturation evidence
+```
+
+---
+
+# 67. Phase 6 — Skill Autoresearch MVP
+
+先只允许 Tier A mutation。
+
+运行：
+
+```text
+baseline
+→ one change
+→ deterministic gates
+→ DEV
+→ keep/reject
+→ log
+```
+
+不做 nightly。
+
+验收：
+
+```text
+protected mutation = INVALID
+bad candidate 自动回退
+results.tsv append-only
+best pointer 正确
+```
+
+---
+
+# 68. Phase 7 — Holdout / Adversarial / Empirical Promotion
+
+实现：
+
+```text
+DEV / HOLDOUT separation
+noise floor
+repeated empirical run
+adversarial suite
+Pareto promotion
+```
+
+验收：
+
+不能通过：
+
+```text
+输出更少
+只在 DEV 特化
+破坏 contradiction recall
+增加巨大复杂度换微小收益
+```
+
+的方式 KEEP。
+
+---
+
+# 69. Phase 8 — Daily Evolution
+
+加入：
+
+```text
+bounded daily profile
+checkpoint
+plateau
+budget
+PR preparation
+```
+
+初版只：
+
+```text
+branch_only
+```
+
+绝不自动 merge。
+
+---
+
+# 70. Phase 9 — UI / Showcase
+
+Research Studio 增加：
+
+```text
+Decision State
+Gap Priority
+Autoresearch Iterations
+Research Saturation
+Revision Diff
+```
+
+Developer 模式增加：
+
+```text
+Skill Evolution
+```
+
+公开报告默认不展示内部自我优化 debug 信息。
+
+---
+
+# 71. 文件变更矩阵
+
+| 文件/目录 | 动作 |
+|---|---|
+| `SKILL.md` | 小改：加入 meta-layer 路由，不塞实现细节 |
+| `docs/architecture.md` | Phase 1 后更新 5 概念分层 |
+| `docs/living-evidence.md` | Phase 4 后加入主动 Gap research loop |
+| `docs/benchmark.md` | Phase 7 更新 DEV/HOLDOUT/ADV |
+| `docs/agent-mcp-enhanced-mode.md` | Phase 1 改成 ExecutionPlanner → TaskSpec → safe_spawn |
+| `skill/agents/*` | 先兼容保留 |
+| `skill/roles/registry.yaml` | 新增 |
+| `engine/gaps.py` | 扩展 priority metadata，不改 coverage truth |
+| `engine/living.py` | 复用 ingestion；增加 controller adapter，不塞 orchestrator |
+| `engine/autoresearch/*` | 新增 |
+| `engine/orchestration/*` | 新增 |
+| `engine/autoevolve/*` | 新增 |
+| `integrations/agent_mcp.py` | Phase 1 改读取 registry / TaskSpec，保留 gate |
+| `benchmarks/*` | 重组 eval partitions，保持历史结果 |
+| `.github/workflows/ci.yml` | 加 scientific/protected gates |
+
+---
+
+# 72. 不做的架构
+
+明确排除：
+
+```text
+“八 Agent 一起跑所有问题”
+“每一个 Stage 一个 Agent”
+“无限 swarm”
+“Agent 互相自由聊天决定研究方向”
+“自动删除不利证据”
+“用一个综合 LLM Judge score 决定科学质量”
+“自动改 evaluator 让 benchmark 变好”
+“自动 merge main”
+“没有 Gap 就自动设计研究”
+“搜不到三轮就声称不存在证据”
+“DVI 伪装成 EVPI 概率”
+```
+
+---
+
+# 73. 旗舰 CS1 案例的最终运行示例
+
+用户问题：
+
+> 大一 C/Python 是否应该允许生成式 AI 编程助手？
+
+初始 Evidence Review：
+
+```text
+Task Performance：有收益
+Independent Learning：冲突
+Transfer：不足
+Retention：不足
+Decision：PILOT
+```
+
+Graph 生成：
+
+```text
+GAP-101 missing_transfer
+GAP-102 missing_retention
+GAP-103 population heterogeneity
+```
+
+DVI：
+
+```text
+GAP-101 HIGH
+GAP-102 MEDIUM
+GAP-103 MEDIUM
+```
+
+Iteration 1：
+
+```text
+Hypothesis：novice CS1 + without-AI transfer targeted query
+Result：找到 1 个 eligible Study
+GraphRevision 8 → 9
+```
+
+Iteration 2：
+
+```text
+重新排名
+GAP-101 仍 HIGH 但 directness 改善
+尝试 citation chaining
+Result：0 eligible, 6 duplicates
+记录 NegativeSearchRecord
+```
+
+Iteration 3：
+
+```text
+独立 counter/risk search
+Result：1 个 relevant Finding
+GraphRevision 9 → 10
+```
+
+Iteration 4：
+
+```text
+search yield 接近 0
+strategy diversity 基本耗尽
+GAP-101 → SEARCH_SATURATED
+```
+
+系统判断：
+
+```text
+HIGH DVI
++ unresolved
++ secondary evidence saturated
+→ EMPIRICAL_EVIDENCE_NEEDED
+```
+
+然后才：
+
+```text
+Gap
+→ StudyDesign
+→ restricted Pilot
+→ Data
+→ New Finding
+→ GraphRevision 11
+→ DecisionSnapshot
+```
+
+这是最终要实现的“研究不是报告终点”的系统行为。
+
+---
+
+# 74. Competition / Product Differentiation
+
+Autoresearch 迭代后，不要宣传：
+
+> “24 小时无人值守自动科研。”
+
+更准确的表达：
+
+> **EduEvidence 不只在新证据出现后更新决定，还会识别当前决策中最值得解决的证据缺口，执行有边界的下一轮研究，并持续用受保护的评估体系改进自己的研究方法。**
+
+英文：
+
+> **EduEvidence does not only update decisions when evidence changes. It identifies which evidence is most worth seeking next, runs bounded research iterations, and continuously tests how to become a better research system without optimizing the conclusion itself.**
+
+---
+
+# 75. 最终 Definition of Done
+
+## Evidence Autoresearch
+
+- [ ] Gap 来源仍是 Evidence Graph，不是自由脑暴
+- [ ] Gap 有透明 DVI band 与 rationale
+- [ ] 每轮一个主 hypothesis
+- [ ] 每轮有明确 budget / termination
+- [ ] Negative search 可追溯
+- [ ] Valid Evidence append-only
+- [ ] No-gain iteration 不生成假 GraphRevision
+- [ ] Search saturation 可解释
+- [ ] 高 DVI + saturated 才能桥接 empirical study
+- [ ] Decision action 不被自动真实执行
+
+## Orchestration
+
+- [ ] Protocol / Role / Capability / Worker / Model 完全分离
+- [ ] S 级默认 0 subagents
+- [ ] M/L 按 evidence axis 并行
+- [ ] TaskSpec 是所有 spawn 的前置条件
+- [ ] safe_spawn 仍是唯一 Agent MCP path
+- [ ] Single Writer
+- [ ] Judge 不读取 unvalidated worker prose
+- [ ] parallel worker cap 生效
+
+## Skill Autoresearch
+
+- [ ] Baseline 必跑
+- [ ] One change per experiment
+- [ ] Protected manifest 生效
+- [ ] DEV / HOLDOUT / ADV 隔离
+- [ ] empirical repeats + noise floor
+- [ ] Pareto promotion
+- [ ] simplicity criterion
+- [ ] results / failures append-only
+- [ ] plateau / budget / ceiling
+- [ ] branch only；不 auto merge main
+
+## Scientific Integrity
+
+- [ ] 不优化 conclusion
+- [ ] 不删除反证
+- [ ] 不伪造 precision
+- [ ] 不把 active-learning ranking 当 evidence weight
+- [ ] 不把 conceptual DVI 当正式 EVSI
+- [ ] 不把 simulation benchmark 当真实性能证据
+- [ ] human-subject Pilot 仍有人类 / ethics gate
+
+---
+
+# 76. 实施优先级
+
+```text
+P0  Orchestration clarity / Single Writer
+↓
+P1  ResearchIteration + NegativeSearch memory
+↓
+P2  DVI / Gap lifecycle
+↓
+P3  bounded Evidence Autoresearch
+↓
+P4  Saturation → Pilot bridge
+↓
+P5  Skill Autoresearch protected MVP
+↓
+P6  Holdout / adversarial / repeated empirical eval
+↓
+P7  Daily evolution
+↓
+P8  Studio visualization
+```
+
+不要先做 UI。
+
+不要先做“每天自动跑”。
+
+先让：
+
+```text
+状态
+责任
+评估
+保护区
+```
+
+四件事正确。
+
+---
+
+# 77. 最终架构一句话
+
+> **EduEvidence vNext = 一个以 Evidence Graph 为事实状态、以 Decision 为目标、以 KnowledgeGap 为下一步研究入口、以 bounded experiments 为学习机制、以 protected evals 约束自我改进、以按需 subagents 扩展执行能力的 Living Decision Research Engine。**