ARIS 对抗式声明审计 (ARIS Adversarial Claim Auditing)

这是什么（一句话） (What This Is — In One Sentence)

给任何深度产出（文章、报告、分析、代码、科研草稿）加一道 "异族 reviewer 挑刺 + 三阶段证据审计" 的兜底闸门，
专门治 ARIS 论文指出的长程任务头号失败模式：看似成立、实则证据不足的结论（plausible unsupported success）。

Add a safety gate of "out-group reviewer critique + three-stage evidence audit" to any deep deliverable (articles, reports, analyses, code, research drafts), specifically targeting the #1 failure mode of long-horizon tasks identified by the ARIS paper: conclusions that look valid but are actually evidence-deficient (plausible unsupported success).

> 它解决的核心问题：模型不是"明显崩了"才出错，而是顺着上下文编出一段听起来很对、但证据链断掉的结论。
> 这种失败肉眼难查，越长的任务越致命。ARIS 的解法是把"验证"做成 harness 的一等公民，而不是靠作者自觉。

> The core problem it solves: models don't fail only by "obviously collapsing"; rather, they spin up a conclusion that sounds right along the context but whose evidence chain is broken.
> This failure is hard to spot by eye and grows more lethal the longer the task. ARIS's solution makes "verification" a first-class citizen of the harness, instead of relying on the author's awareness.

什么时候用 (When To Use)

满足任意一条就用：

Use it whenever any one of the following applies:

• 要发长文/技术报告/公众号文章，担心结论站不住、被读者挑刺
• 做了分析/研究，怕"中间推导太顺、最后结论飘了"
• 产出要可被追溯（每个结论都能点名证据）

• 用户说"帮我审一下这篇""这个结论有依据吗""按 ARIS 那种对抗审查来一遍"

• Publishing a long-form article / technical report / WeChat public-account piece and worried the conclusions won't hold up or readers will poke holes.

• Did an analysis / research and afraid the "middle reasoning went too smoothly, but the final conclusion drifted."
• The deliverable must be traceable (every conclusion can cite its evidence).
• The user says "help me audit this," "does this conclusion have a basis," or "run that adversarial review like ARIS does."

不适用：

Not applicable:

• 一步能答完的简单问答、纯闲聊
• 纯创作（小说/诗歌）不需要证据链

• 已经有更强的人工同行评审流程且不缺这一步

• Simple Q&A answerable in one step, or pure small talk.

• Pure creative writing (novels/poetry) that needs no evidence chain.
• Already has a stronger human peer-review process and doesn't need this step.

受众（谁该用） Audience (Who Should Use)

面向所有 Agent 用户，从新手到专业：

For all Agent users, from novice to professional:

• 新手：照 references/quickstart-minimal-example.md 三步跑通脚本，先建立"每个结论都要有证据"的直觉——不必懂对抗审查也能用。
• 进阶：用异族 reviewer 子 agent 跑完整对抗审查，处理复杂格式与边界。

• 专业：把本技能嵌入团队发布流程，配合 wiki/research/ 做跨次经验沉淀与自改进。

• Novice: run the 3-step script in references/quickstart-minimal-example.md to build the intuition that "every claim needs evidence" — usable without understanding adversarial review.

• Intermediate: run the full adversarial review with an out-group reviewer sub-agent, handling complex formats and edge cases.
• Professional: embed this skill into a team's publishing pipeline, pairing with wiki/research/ for cross-run experience persistence and self-improvement.

自动触发条件（Auto-trigger Conditions）

本技能作为 WorkBuddy 技能，在以下任一条件满足时应自动激活（无需手动点名）：

This skill, as a WorkBuddy skill, should auto-activate when any one of the following holds (no manual invocation needed):

• 显式触发：用户说出触发词（声明审计 / 对抗审查 / 证据核查 / ARIS / claim audit / 结论有没有证据 / 长文质检）。
• 长文 + 断言信号：待审产物 ≥ ~500 字，且含断言标记（比 / 优于 / 证明 / 表明 / 准确率 / 倍 / 显著 / 领先 / 必然 …）。

• 高风险发布：产物即将对外发布（公众号 / 报告 / 官网），且含数字结论或对比结论。

• Explicit trigger: the user says a trigger word (声明审计 / 对抗审查 / 证据核查 / ARIS / claim audit / 结论有没有证据 / 长文质检).

• Long-form + claim signal: the artifact is ≥ ~500 words and contains claim markers (比 / 优于 / 证明 / 表明 / 准确率 / 倍 / 显著 / 领先 / 必然 …).
• High-risk publish: the artifact is about to be published externally (public account / report / website) and contains numeric or comparative conclusions.

不满足条件时不要强行套用；简单问答、纯闲聊、纯创作无需审计。
When none of these hold, do not force-apply; simple Q&A, pure small talk, and pure creative writing need no audit.

CHECKPOINT 0 — 启动前确认 (Pre-launch Confirmation) (checkpoint_design)

显式标记：在执行任何对抗审查前，必须输出并核对以下启动状态，作为 integrity_checklist.md 的"阶段 1"前置检查：

Explicit marker: Before running any adversarial review, you must output and verify the following launch status as the "Stage 1" pre-check of integrity_checklist.md:

• [ ] CK-0.1 指定了"异族" reviewer（或至少独立子 agent）
• [ ] CK-0.2 明确了最大对抗轮数（默认 3 轮）
• [ ] CK-0.3 指定了产物目录，确保 claim_ledger.md、claim_audit_report.md、integrity_checklist.md 可落地

• [ ] CK-0.4 确认待审产物已包含中间证据（原始日志/文件/截图路径可追溯）

• [ ] CK-0.1 Designated an "out-group" reviewer (or at least an independent sub-agent).

• [ ] CK-0.2 Clarified the maximum number of adversarial rounds (default 3 rounds).
• [ ] CK-0.3 Designated the artifact directory, ensuring claim_ledger.md, claim_audit_report.md, integrity_checklist.md can be produced.
• [ ] CK-0.4 Confirmed the artifact under review already contains intermediate evidence (raw logs/files/screenshot paths traceable).

核心流程 A：对抗式编排（executor ↔ reviewer） (Core Flow A: Adversarial Orchestration)

ARIS 的默认配置是跨模型家族对抗——executor 用一个模型推进，reviewer 用另一个家族的模型审阅并强制要求修改。
异族比同族更能跳出"同一个训练偏差"。

ARIS's default configuration is cross-model-family adversarial — the executor advances with one model, and the reviewer reviews and forcibly demands revisions using a model from another family. An out-group reviewer escapes "the same training bias" better than a same-family one.

角色分工 (Role Division)

Reviewer 审查 Prompt 模板（直接套） (Reviewer Review Prompt Template — Use Directly)

你是无利益相关的独立评审。下面是 executor 的产物（及它的中间证据）。
请只做批判性审查，不要重写，输出结构化意见：

1. 未证断言（Unsupported claims）：哪些结论缺少直接证据？列出原文句子 + 缺的证据类型。
2. 逻辑跳跃（Logical gaps）：从证据 A 到结论 B 之间缺了哪一步推理？
3. 静默继承（Silent inheritance）：有没有把上游/前一步的假设未经证实就当成事实接进来？
4. 反例/边界（Counterexamples）：这个结论在什么条件下会不成立？
5. 修订要求（Mandatory revisions）：按严重程度 P0/P1/P2 列出 executor 必须改的点。

规则：不客气、不凑数；每条意见必须点名具体原文位置；P0 不修就打回。

CHECKPOINT 1 — 对抗循环纪律 (Adversarial Loop Discipline) (checkpoint_design)

显式标记：每轮对抗迭代前后必须输出，作为 claim_audit_report.md 的"过程可追溯"证据。

Explicit marker: Must be output before and after each adversarial iteration, as "process traceability" evidence for claim_audit_report.md.

• [ ] CK-1.1 每轮 executor 改完，必须回指 reviewer 的具体意见编号再改，禁止"假装改了"
• [ ] CK-1.2 设最大轮数 3 轮；第 3 轮 reviewer 仍打回 P0 → 收口并标注"未达成共识的争议点"
• [ ] CK-1.3 reviewer 不得替 executor 写内容，只审只挑——避免变成两个 executor 互相包庇

• [ ] CK-1.4 每轮结束后更新 references/results.tsv：记录轮次、意见编号、闭合状态（已闭合/待确认/争议收口）

• [ ] CK-1.1 After the executor finishes revisions each round, it must cite the reviewer's specific comment number before revising; "pretending to fix" is forbidden.

• [ ] CK-1.2 Set a max of 3 rounds; if the reviewer still rejects a P0 in round 3 → close out and mark "points of unresolved disagreement."
• [ ] CK-1.3 The reviewer must not write content for the executor; only reviews and critiques — to avoid two executors covering for each other.
• [ ] CK-1.4 After each round, update references/results.tsv: record the round, comment number, and closure status (closed / pending confirmation / disputed close-out).

隐含步骤补充（规则5修复） (Implicit-Step Supplements — Rule 5 Fix)

reviewer 意见编号规则：每轮 reviewer 输出用 R{轮次}.{序号} 编号（如 R1.1 / R1.2 / R2.1），executor 修订时回指具体编号，禁止模糊回指。

Reviewer comment numbering rule: Each round's reviewer output is numbered R{round}.{seq} (e.g., R1.1 / R1.2 / R2.1); when the executor revises, it must cite the specific number; vague references are forbidden.

中间证据引用格式：证据来源统一写为 文件名:行号 或 截图/日志文件路径 + 时间戳，例如 exp/benchmark_0421.md:L12、reviewer_output_round1.txt:L5-8。

Intermediate evidence citation format: Evidence sources are uniformly written as filename:linenumber or screenshot/log file path + timestamp, e.g., exp/benchmark_0421.md:L12, reviewer_output_round1.txt:L5-8.

references/results.tsv 最小列结构（8列，制表符分隔）：

Minimum column structure of references/results.tsv (8 columns, tab-separated):

timestamp    case    input_summary    artifacts    reviewer_verdict    rounds    status    note

示例：

Example:

2026-07-22T14:00+08:00    T1    "本地部署DeerFlow比云端快3倍"    references/claim_ledger.md    P0-closed    1    passed    降级路径:L1

无法异族审核时的记录要求：在 results.tsv 的 note 列标注降级层级（L1/L2/L3）及原因。

Recording requirement when out-group review is impossible: In the note column of results.tsv, mark the degradation level (L1/L2/L3) and the reason.

核心流程 B：三阶段声明审计（保证层） (Core Flow B: Three-Stage Claim Audit — Assurance Layer)

这是 ARIS 保证层的精华，把"结论有没有证据"拆成三段可执行的检查。

This is the essence of ARIS's assurance layer, splitting "does the conclusion have evidence" into three executable checks.

阶段 1 — 完整性验证（Integrity Verification） (Stage 1 — Integrity Verification)

• 核对每个实验结果/数据是否可复现、来源完整（原始日志/文件/截图是否齐）
• 标记"结果缺失 / 被截断 / 只报好看的部分"的段落

• 产物：integrity_checklist.md（每项打 ✅/⚠️/❌）

• Verify each experimental result/data is reproducible and complete in source (whether raw logs/files/screenshots are present).

• Flag paragraphs that are "results missing / truncated / only reporting the good parts."
• Artifact: integrity_checklist.md (each item marked ✅/⚠️/❌).

CHECKPOINT 2 — 完整性验证完成标记 (Integrity Verification Completion Marker) (checkpoint_design)

显式标记：阶段 1 结束时必须输出以下标记，进入阶段 2 前不得跳步：

Explicit marker: At the end of Stage 1, the following marker must be output; do not skip ahead to Stage 2:

• [ ] CK-2.1 integrity_checklist.md 已生成，所有实验/数据项已标注 ✅/⚠️/❌
• [ ] CK-2.2 标记为 ❌ 或 ⚠️ 的项目已在 references/results.tsv 登记为"待补证据"

• [ ] CK-2.3 若有关键证据缺失，先补证据或降级措辞，不允许直接放行

• [ ] CK-2.1 integrity_checklist.md generated; all experiment/data items marked ✅/⚠️/❌.

• [ ] CK-2.2 Items marked ❌ or ⚠️ are registered in references/results.tsv as "evidence to be supplemented."
• [ ] CK-2.3 If critical evidence is missing, supplement evidence or downgrade wording first; direct pass-through is not allowed.

阶段 2 — 结果→声明映射（Result-to-Claim Mapping） (Stage 2 — Result-to-Claim Mapping)

• 建一张 Claim Ledger（声明台账），强制"每个声明↔它的证据行"

• 模板：

• Build a Claim Ledger, enforcing "each claim ↔ its evidence row."

• Template:

| # | 声明(原文句子) | 证据来源(文件/行/截图) | 支撑强度 | 缺口 |
|---|---|---|---|---|
| C1 | "X 比 Y 快 2 倍" | exp/log_0421.txt L12 | 强 | 无 |
| C2 | "方案适用于生产" | （无直接实验） | 弱 | 缺压测数据 |

• 任何填不进台账的声明 = 红色缺口，必须要么补证据、要么降级措辞、要么删

• Any claim that can't fit into the ledger = a red gap; it must either be supplemented with evidence, downgraded in wording, or deleted.

CHECKPOINT 3 — 声明台账完成标记 (Claim Ledger Completion Marker) (checkpoint_design)

显式标记：阶段 2 结束时必须输出：

Explicit marker: At the end of Stage 2, the following must be output:

• [ ] CK-3.1 claim_ledger.md 已生成，成稿中的每个陈述都有台账行或显式缺口标记
• [ ] CK-3.2 红色缺口（无证据的声明）已标注为"必须补证据 / 降级 / 删除"三选一

• [ ] CK-3.3 references/results.tsv 已更新：新增 claim_ledger.md 的链接与行数

• [ ] CK-3.1 claim_ledger.md generated; every statement in the draft has a ledger row or explicit gap marker.

• [ ] CK-3.2 Red gaps (claims without evidence) are marked as one of "must supplement evidence / downgrade / delete."
• [ ] CK-3.3 references/results.tsv updated: added link and line count for claim_ledger.md.

阶段 3 — 声明审计（Claim Auditing） (Stage 3 — Claim Auditing)

• 把成稿里的每个陈述和 Claim Ledger 交叉比对：
• 成稿说了但台账没有 → 幻觉/漂移，必须修
• 台账有但成稿没说 → 可补回或标注为"未采用"

• 产物：claim_audit_report.md（缺口清单 + 必须改项 + 建议改项）

• Cross-check every statement in the draft against the Claim Ledger:

• Stated in draft but absent from ledger → hallucination/drift, must fix.
• In ledger but not stated in draft → can be added back or marked "not adopted."
• Artifact: claim_audit_report.md (gap list + must-fix items + suggested-fix items).

CHECKPOINT 4 — 声明审计完成标记 (Claim Audit Completion Marker) (checkpoint_design)

显式标记：阶段 3 结束时必须输出：

Explicit marker: At the end of Stage 3, the following must be output:

• [ ] CK-4.1 claim_audit_report.md 已生成，包含缺口清单、必须改项、建议改项
• [ ] CK-4.2 报告中链接了 references/results.tsv 与 references/evolution-log.md 作为过程记录
• [ ] CK-4.3 所有 P0 打回项已闭合，未闭合的标注为"待用户确认"

• [ ] CK-4.4 运行 python scripts/aris_audit.py gate --ledger claim_ledger.md 必须通过（返回 0） 才允许 Finish；返回非零即说明台账仍有未验证声明，禁止收口

• [ ] CK-4.1 claim_audit_report.md generated, containing gap list, must-fix items, suggested-fix items.

• [ ] CK-4.2 The report links references/results.tsv and references/evolution-log.md as process records.
• [ ] CK-4.3 All P0 rejected items are closed; unclosed ones marked "pending user confirmation."
• [ ] CK-4.4 Running python scripts/aris_audit.py gate --ledger claim_ledger.md must pass (return 0) before Finish; a non-zero return means the ledger still has unverified claims — closing is forbidden.

> 三阶段顺序不可跳：先保证"证据本身没假"，再保证"声明都 mapped"，最后保证"成稿和台账一致"。
> 每完成一个阶段，必须输出对应 CHECKPOINT（CK-2 / CK-3 / CK-4）并在 references/results.tsv 留痕。

> The three-stage order cannot be skipped: first ensure "the evidence itself isn't fake," then ensure "all claims are mapped," finally ensure "the draft and ledger are consistent."
> After completing each stage, you must output the corresponding CHECKPOINT (CK-2 / CK-3 / CK-4) and leave a trace in references/results.tsv.

references/ 目录结构（directory_structure，gap2 修复） (references/ Directory Structure — directory_structure, gap2 Fix)

为保证审计过程可追溯、可回归，每个审计实例建议落地以下目录结构。所有文件均为相对路径，可在任意 Hermes 环境直接使用。

To ensure the audit process is traceable and reproducible, each audit instance is recommended to use the following directory structure. All files are relative paths and can be used directly in any Hermes environment.

<审计工作目录>/
├── SKILL.md                                    # 本技能定义
├── scripts/
│   ├── aris_audit.py                           # 前置自审脚本（阶段 2/3）
│   └── test_aris_audit.py                      # 回归测试（23 项）
└── references/
    ├── README.md                               # 本目录说明与文件清单
    ├── evolution-log.md                        # 演进日志（本次 v1.3.1 的变更与决策）
    ├── results.tsv                             # 实测结果登记表（T1/T2 及后续用例）
    ├── integrity_checklist.md                  # 阶段 1 产物
    ├── claim_ledger.md                         # 阶段 2 产物
    └── claim_audit_report.md                   # 阶段 3 产物（审计报告，链接至 results.tsv）

> references/ 目录不承载任何运行时逻辑，仅作为审计证据与演进记录的容器。脚本默认输出到当前工作目录，审计者应手工或按 Runbook 将产物归档到 references/。

> The references/ directory carries no runtime logic; it only serves as a container for audit evidence and evolution records. Scripts output to the current working directory by default; the auditor should manually or per the Runbook archive the artifacts into references/.

核心流程 C：持久化 research wiki + 自改进循环（已落地） (Core Flow C: Persistent Research Wiki + Self-Improvement Loop — Implemented)

ARIS 执行层的"持久化 research wiki"在本项目直接落地为 LLM Wiki（图书馆 vault）——长程任务要的是知识沉淀层，不是只靠上下文窗口。配套 wiki-lint 做健康巡检。

ARIS's execution-layer "persistent research wiki" is directly implemented in this project as the LLM Wiki (Library vault) — long-horizon tasks need a knowledge-persistence layer, not just the context window. Paired with wiki-lint for health checks.

C.0 仓库与目录约定 (C.0 Repo & Directory Conventions)

• 持久化 wiki 根：D:\图书馆\wiki\（编译层）。可在任意环境重写为任意 LLM Wiki 路径（保持可移植）。
• ARIS 审计产物目录 wiki/research/（类型 aris-audit / aris-pitfall / aris-log）：
• aris-pitfall-register.md — 跨实例坑点登记表（自改进循环的"记忆"）
• aris-research-log.md — 审计实例登记表（每次运行追加一行）
• aris-audit-{slug}.md — 单次审计的持久化实例（声明台账摘要 / 被推翻假设 / reviewer 共识 / verdict）

• 分工：运行态 references/（skill 内）承载当次工作产物（integrity_checklist / claim_ledger / claim_audit_report / evolution-log / results.tsv）；知识态 wiki/research/ 承载跨次复用的沉淀。两者不混。

• Persistent wiki root: D:\图书馆\wiki\ (compilation layer). Can be rewritten to any LLM Wiki path in any environment (keeping it portable).

• ARIS audit artifact directory wiki/research/ (types aris-audit / aris-pitfall / aris-log):
• aris-pitfall-register.md — cross-instance pitfall register (the "memory" of the self-improvement loop).
• aris-research-log.md — audit instance register (append one line per run).
• aris-audit-{slug}.md — persistent instance of a single audit (claim ledger summary / overturned assumptions / reviewer consensus / verdict).
• Division: Runtime state references/ (within the skill) holds the current run's work artifacts (integrity_checklist / claim_ledger / claim_audit_report / evolution-log / results.tsv); knowledge state wiki/research/ holds cross-run reusable deposits. The two do not mix.

C.1 读前引导（Bootstrap · 复用历史发现） (C.1 Bootstrap — Reuse Historical Findings)

每次审计启动（CHECKPOINT 0 之前），Executor 必须先读：

Before each audit launch (before CHECKPOINT 0), the Executor must first read:

• wiki/research/aris-pitfall-register.md —— 规避已知坑（如"异族降级须登记""证据缺口必须显式降级"）
• wiki/research/aris-research-log.md + 相关 aris-audit-*.md —— 继承 reviewer 共识、复用可迁移结论

• 若环境无 wiki（纯技能自包含模式），跳过本步并在 references/results.tsv 标注 wiki=offline。

• wiki/research/aris-pitfall-register.md — avoid known pitfalls (e.g., "out-group downgrade must be registered," "evidence gaps must be explicitly downgraded").

• wiki/research/aris-research-log.md + relevant aris-audit-*.md — inherit reviewer consensus, reuse transferable conclusions.
• If the environment has no wiki (pure self-contained skill mode), skip this step and mark wiki=offline in references/results.tsv.

C.2 写回（Write-back · 沉淀本次发现） (C.2 Write-back — Persist This Run's Findings)

审计收口（所有 P0 闭合）后，Executor 写回：

After the audit closes (all P0 closed), the Executor writes back:

• 新建/更新 wiki/research/aris-audit-{slug}.md：声明台账摘要表 + 被推翻的假设 + reviewer 共识 + 最终 verdict。
• 在 wiki/research/aris-research-log.md 追加一行：日期 | 实例 | 涉及点(1/2/3) | reviewer verdict | harness 改进 | 链接。

• 若本次浮现新的反复坑，更新 aris-pitfall-register.md（新增一行 + 详情）。

• Create/update wiki/research/aris-audit-{slug}.md: claim ledger summary table + overturned assumptions + reviewer consensus + final verdict.

• Append one line to wiki/research/aris-research-log.md: date | instance | points involved (1/2/3) | reviewer verdict | harness improvement | link.
• If a new recurring pitfall emerges this run, update aris-pitfall-register.md (add a line + details).

C.3 自改进闸门（Reviewer-Approval Gate · 防正反馈失控） (C.3 Self-Improvement Gate — Reviewer-Approval Gate)

• Executor 可提出 harness 改进（新 CHECKPOINT / 新 anti-pattern / 脚本规则），但只能写成"提案"写入 references/evolution-log.md，状态标 待评审。
• 必须经异族 reviewer 批准才合入 SKILL.md——禁止"自己改自己"。
• reviewer 批准后，状态改 已采纳 并合入；驳回则 被驳回 留痕。references/results.tsv 记录采纳状态。

• 触发 wiki-lint（只读巡检）确保新页面无断链 / 孤立；索引同步更新。

• The Executor may propose harness improvements (new CHECKPOINT / new anti-pattern / script rule), but only as a "proposal" written to references/evolution-log.md with status 待评审 (pending review).

• Must be approved by an out-group reviewer before merging into SKILL.md — "editing oneself" is forbidden.
• After reviewer approval, status changes to 已采纳 (adopted) and merges; if rejected, 被驳回 (rejected) is recorded. references/results.tsv records the adoption status.
• Trigger wiki-lint (read-only check) to ensure new pages have no broken links / orphans; the index syncs.

> 闭环口诀：读前复用 → 跑 1+2 → 写回沉淀 → 提案过闸。知识态与运行态分离，reviewer 闸门保证自改进不跑偏。

> Closed-loop mantra: reuse before reading → run 1+2 → write back to persist → proposal passes the gate. Knowledge state and runtime state are separated; the reviewer gate ensures self-improvement doesn't go off track.

实操 Runbook（怎么真跑起来） (Practical Runbook — How To Actually Run It)

标准路径（支持子 agent + 异族模型） (Standard Path — Sub-agent + Out-group Model Supported)

1. 主 agent = Executor：正常推进任务、产出草稿。
2. 派 Reviewer 子 agent：用 Agent 工具 spawn 一个独立子 agent，prompt 里贴"Reviewer 审查模板" + 当前产物 + 证据文件；若环境支持，把子 agent 的 model 设为与主 agent 不同家族的模型（如主用闭源、reviewer 用开源）。
3. 三阶段审计：主 agent 按阶段 1→2→3 跑，产出 claim_ledger.md 与 claim_audit_report.md；每阶段结束时输出对应 CHECKPOINT。
4. 收口闸门：reviewer 打回的 P0 全部闭合后，运行 python scripts/aris_audit.py gate --ledger claim_ledger.md；返回 0 才允许 Finish，返回非零说明台账仍有未验证声明，禁止收口。未闭合的标注为"待用户确认"。

5. 归档：将 integrity_checklist.md、claim_ledger.md、claim_audit_report.md 移入 references/，并在 references/results.tsv 登记。

6. Main agent = Executor: advance the task normally and produce a draft.

7. Spawn a Reviewer sub-agent: use the Agent tool to spawn an independent sub-agent, pasting the "Reviewer review template" + current artifacts + evidence files into the prompt; if the environment supports it, set the sub-agent's model to a model from a different family than the main agent (e.g., main uses closed-source, reviewer uses open-source).
8. Three-stage audit: the main agent runs stages 1→2→3, producing claim_ledger.md and claim_audit_report.md; output the corresponding CHECKPOINT at the end of each stage.
9. Closure gate: after all P0 rejected by the reviewer are closed, run python scripts/aris_audit.py gate --ledger claim_ledger.md; only allow Finish on return 0; a non-zero return means the ledger still has unverified claims — closing is forbidden. Unclosed ones marked "pending user confirmation."
10. Archive: move integrity_checklist.md, claim_ledger.md, claim_audit_report.md into references/, and register in references/results.tsv.

Executor(主) ──产出──> Reviewer(子,异族) ──挑刺──> Executor 修订 ──> 循环(≤3)
                                              │
                              +── 三阶段声明审计(主 agent 跑) ──> Claim Ledger + Audit Report

步数限制与兜底（防失控） Step Budget & Fallback (Anti-runaway)

为应对"长文 / 复杂场景缺少保护机制"的可靠性短板，本技能对整体过程设硬上限（兜底保障，宁可标注争议、不可无限空转）：

To address the reliability gap of "long/complex runs lacking protection mechanisms," this skill enforces hard caps on the overall process (fallback guarantee: better to mark disagreement than loop forever):

• 对抗轮数 ≤ 3：第 3 轮 reviewer 仍打回 P0 → 收口并标注"未达成共识的争议点"，禁止死循环。
• 单轮修订预算 ≤ 5 处：每轮 executor 只聚焦 reviewer 点名的修订点，避免范围无限扩张。
• 整体步数预算 ≤ 30 步：从 CHECKPOINT 0 到 Finish 的 agent 步数上限；超预算即收口，未决项标注"待用户确认"。

• 硬闸门 gate：收口前必须 gate 通过（见 CHECKPOINT 4 / CK-4.4），否则机制性阻止 Finish。

• Adversarial rounds ≤ 3: if round 3 still rejects a P0 → close out and mark "points of unresolved disagreement"; dead loops forbidden.

• Per-round revision budget ≤ 5: each round the executor focuses only on the reviewer's cited fixes, avoiding unbounded scope expansion.
• Overall step budget ≤ 30: agent step cap from CHECKPOINT 0 to Finish; over budget → close out, unresolved items marked "pending user confirmation."
• Hard gate: gate must pass before closing (see CHECKPOINT 4 / CK-4.4), else mechanically blocks Finish.

> 所有未决 / 争议项必须在 claim_audit_report.md 显式列出，不得静默吞掉。
> All unresolved / disputed items must be explicitly listed in claim_audit_report.md; never silently dropped.

> 想先看一遍真实跑通效果？见 references/quickstart-minimal-example.md（自带示例文件，输出均为脚本真实产物）。审计报告成品长什么样？见 references/sample-claim-audit-report.md。
> Want to see a real run first? See references/quickstart-minimal-example.md (bundled examples, outputs are real script output). What a finished audit report looks like? See references/sample-claim-audit-report.md.

降级路径（不支持子 agent / 异族模型时） (Degradation Path — When Sub-agent / Out-group Model Unsupported)

若当前环境无法 spawn 子 agent 或配置异族模型，按以下顺序降级，并在 references/results.tsv 标注降级原因：

If the current environment cannot spawn a sub-agent or configure an out-group model, degrade in the following order, and mark the degradation reason in references/results.tsv:

> 降级路径不改变三阶段声明审计的核心纪律（阶段顺序不可跳、P0 必改、CHECKPOINT 必输出），只改变 reviewer 的实现方式。

> The degradation path does not change the core discipline of the three-stage claim audit (stage order cannot be skipped, P0 must be fixed, CHECKPOINT must be output); it only changes how the reviewer is implemented.

脚本化前置自审（aris_audit.py） (Scripted Pre-Audit — aris_audit.py)

配套脚本 scripts/aris_audit.py（单文件、零依赖、四子命令）把阶段 2/3 里最机械、最该先做的部分自动化，
让你在派 Hermes + SenseNova 异族审之前，先自己降一轮缺口。

The companion script scripts/aris_audit.py (single-file, zero-dependency, four subcommands) automates the most mechanical, should-do-first parts of stages 2/3, letting you reduce a round of gaps yourself before dispatching the Hermes + SenseNova out-group review.

> 定位（务必记住）：这是前置自审、降缺口用的"粗滤网"，不替代 Hermes + SenseNova 异族审。
> 脚本是纯规则启发式（默认 light 模式，不调 LLM、不联网、零交互），会漏判也会误判，
> 它的产出只是"待人工/异族复核的候选清单"，最终质量闸门仍是异族对抗审查。

> Positioning (must remember): This is a "coarse filter" for pre-audit and gap reduction, not a replacement for the Hermes + SenseNova out-group review.
> The script is pure rule-based heuristics (default light mode, no LLM calls, no network, zero interaction); it will both miss and misjudge,
> its output is only a "candidate list pending human/out-group review"; the final quality gate remains the out-group adversarial review.

设计约束（已固化进脚本） (Design Constraints — Baked Into the Script)

• 零交互：所有选项走 flag，无 input() 询问；默认 light 模式。
• 证据缺口留空标 ⚠️，不追问：extract 生成的台账证据列一律 ⚠️，绝不卡你补证据。
• 单文件四子命令：extract（阶段 2）+ drift（阶段 3）+ gate（收口硬闸门）+ reconcile（跨文档证据核验）。gate 在收口前扫描台账，只要还有未验证声明（证据列仍 ⚠️ 且未显式降级/删除）就返回非零退出码，从机制上阻止"带缺口发稿"。reconcile 用来源文档兜底核验"无来源支撑的断言"。
• 复杂格式识别：extract/drift 已支持 Markdown 表格单元格、加粗（**x**）、项目符号、标题后声明的抽取，避免"复杂格式识别不全"导致的漏检。

• 脚本覆盖范围：脚本仅自动化阶段 2（草稿→台账骨架）和阶段 3（漂移检测）。阶段 1（完整性验证：文件是否齐全、口径是否统一、关键证据是否到位）是人工步骤，在 Runbook ③ 中由人工或异族审完成，脚本不替代。（v2.0.0 起，reconcile 子命令提供跨文档证据核验，作为阶段 2/3 之外的「异族证据兜底」——专门抓"无来源支撑"的断言，仍不替代阶段 1 人工完整性验证。）

• Zero interaction: all options go through flags, no input() prompts; default light mode.

• Leave evidence gaps empty marked ⚠️, don't ask: the ledger evidence column generated by extract is always ⚠️, never blocking you from supplying evidence.
• Single-file four subcommands: extract (stage 2) + drift (stage 3) + gate (stage-4 closure hard-gate) + reconcile (cross-document evidence check).
• Complex-format recognition: extract/drift now also catch claims inside Markdown table cells, bold (**x**), bullet points, and post-heading lines, avoiding missed detections from "incomplete complex-format recognition."
• Script coverage: the script automates stage 2 (draft → ledger skeleton), stage 3 (drift detection), and the stage-4 closure hard-gate (gate). Stage 1 (integrity verification: whether files are complete, caliber is consistent, key evidence is in place) is a manual step, completed by a human or out-group reviewer in Runbook ③; the script does not replace it.

测试与回归验证 (Testing & Regression Validation)

配套回归测试 scripts/test_aris_audit.py（23 项，用固定 fixture 跑 extract/drift/gate/reconcile，断言台账含 ⚠️、父目录自动创建、非 UTF-8 干净报错、未闭环台账 gate 非零退出、漂移类型 B 弱声明被强化、标题后声明检测、表格单元格/加粗声明识别、--json 输出、目录/二进制/空/超大文件干净报错（退出码 2/3/4/7）、gate 退出码 10、reconcile 跨文档证据核验（有/无来源支撑分检 + 严重度 P0/P1/P2 分级）、extract --json 带 severity 分级等）。运行：

The companion regression test scripts/test_aris_audit.py (23 items, using fixed fixtures to run extract/drift/gate/reconcile, asserting the ledger contains ⚠️, parent directory auto-created, clean error on non-UTF-8, gate non-zero exit on unclosed ledger, drift type B weak claim strengthened, post-heading claim detection, table-cell/bold claim recognition, --json output, clean errors on directory/binary/empty/oversized files (exit codes 2/3/4/7), gate exit code 10, reconcile cross-document check (supported/unsupported split + severity P0/P1/P2 tiers), extract --json severity tier, etc.). Run:

python scripts/test_aris_audit.py
# 预期：23/23 passed

extract（阶段 2）：草稿 → 台账骨架 (extract — Stage 2: Draft → Ledger Skeleton)

python scripts/aris_audit.py extract --draft 草稿.md --output ledger.md
# 证据列留空标 ⚠️，候选声明靠规则（含数字/断言词/表格单元格/加粗）抽取，人工再润色
# 可选项：--json（输出 JSON 数组，便于 CI）   --mark ✗（自定义缺口标记，须与 gate 一致）

产出 ledger.md：每张候选声明一行，证据来源 / 支撑强度 / 缺口 三列全留 ⚠️，等你填。

Produces ledger.md: one line per candidate claim, the three columns evidence source / support strength / gap all left as ⚠️, for you to fill.

drift（阶段 3）：成稿 + ledger → 漂移清单 (drift — Stage 3: Draft + Ledger → Drift List)

python scripts/aris_audit.py drift --draft 成稿.md --ledger ledger.md --output drift.md
# 可选项：--json（输出 JSON 对象）   --threshold 0.6（成稿与台账声明相似度阈值，默认 0.6）

输出两类疑似漂移（均标 ⚠️，待复核）：

Outputs two types of suspected drift (both marked ⚠️, pending review):

• A. 成稿有、台账无：成稿里的声明在台账里找不到对应 → 疑似幻觉/漂移。

• B. 弱声明被强化：台账条目仍标 ⚠️（未验证）或标"弱"，但成稿措辞更绝对（出现"最/绝对/远超"等强化词）→ 未验证就被说死。

• A. In draft, not in ledger: a claim in the draft has no corresponding entry in the ledger → suspected hallucination/drift.

• B. Weak claim strengthened: the ledger entry is still marked ⚠️ (unverified) or "weak," but the draft wording is more absolute (using strengthening words like "most/absolute/far exceeds") → stated as settled without verification.

reconcile（v2.0.0 跨文档证据核验）：草稿 + 来源 → 不支持声明清单 (reconcile — v2.0.0 cross-document evidence check)

python scripts/aris_audit.py reconcile --draft 成稿.md --source 论文.md 数据.csv.md --output reconcile.md
# 可选项：--json（输出 JSON 对象）   --threshold 0.3（声明与来源句子的 Jaccard 相似度阈值，默认 0.3）

输出跨文档证据核验报告：
Outputs a cross-document evidence report:

• 来源覆盖率：草稿声明中能在来源文档里找到支撑（token Jaccard ≥ 阈值）的比例。
• 不支持声明清单：覆盖率低于阈值的声明，按严重度优先排列：
• P0：含数字/百分比/倍数的量化声明无来源 → 最高风险（最易被"看似成立"）。
• P1：含"证明/必然/所有/一定/完全"等强绝对词的声明无来源 → 高风险。
• P2：其他未验证声明 → 中低风险。

> 这是 extract/drift/gate 之外的「异族证据兜底」：即便台账自审通过，仍可能漏掉"无来源支撑"的断言，reconcile 专门抓这一类。覆盖度为轻量启发式（声明与来源句子的 token Jaccard），相似度低仅提示"来源中可能无直接支撑"，不证明为假；最终以人工/异族审为准。本工具不替代 Hermes+SenseNova 异族审。

> This is the "out-group evidence backstop" beyond extract/drift/gate: even if the ledger self-audit passes, claims with no source support can slip through; reconcile specifically catches these. Coverage is a lightweight heuristic (token Jaccard between claim and source sentences); low similarity only hints "source may lack direct support," not proof of falsehood; final call rests with human/out-group review.

可选项与错误处理（新手友好） (Options & Error Handling — Newbie-Friendly)

四个子命令共享 --max-bytes（文件大小上限，默认 5MB，超限报错 ERR-TOO-LARGE）。其余可选 flag：

错误处理（统一错误码 + 退出码 + 一行修复提示）：所有可预期错误都带 ERR-* 错误码并打印修复建议，不再出现"专业抽象看不懂"的报错。

Options & Error Handling — Newbie-Friendly

All four subcommands share --max-bytes (file size cap, default 5MB; exceeds → ERR-TOO-LARGE). Other optional flags:

Error handling (unified error code + exit code + one-line fix): every expected error carries an ERR-* code and prints a fix hint — no more "abstract, unreadable" errors.

衔接 Runbook 的推荐顺序 (Recommended Order Linking to the Runbook)

① 写草稿 → ② extract 出 ledger（证据全 ⚠️）
③ 人工/异族填证据、定强弱 → ④ 写定稿
⑤ drift 扫一遍：把"成稿有台账无 / 弱声明被强化"先降一轮
⑥ 再派 Hermes + SenseNova 异族审做最终质量闸门
⑦ 归档至 references/，并在 results.tsv 登记

脚本把 ②③④⑤ 的机械活儿先干了，异族审只需聚焦脚本捞不出的深层逻辑问题。

The script does the mechanical work of ②③④⑤ first; the out-group review only needs to focus on deep logical issues the script can't catch.

测试用例与预期判据（实测表现 gap1 修复） (Test Cases & Expected Criteria — gap1 Fix)

> 本节为 测试用例 + 预期判据，用于验证本技能在关键失败场景下的行为是否正确。
> T1/T2 是人工实测用例（需要 executor+reviewer 子 agent 协作完成，无法在静态测试中自动化）。
> 脚本级回归测试见 scripts/test_aris_audit.py（23 项，覆盖 extract/drift/gate/reconcile/父目录创建/编码错误/未闭环 gate 非零退出(10)/漂移类型B/标题后声明检测/表格单元格声明/加粗声明/--json 输出/目录·二进制·空·超大文件干净报错(2/3/4/7)/--mark 自定义标记/extract --json 严重度分级/reconcile 跨文档证据核验与 P0/P1/P2 分级等）。

> This section is test cases + expected criteria, used to verify whether this skill behaves correctly under key failure scenarios.
> T1/T2 are human-tested cases (require executor+reviewer sub-agent collaboration, cannot be automated in static tests).
> For script-level regression tests, see scripts/test_aris_audit.py (23 items, covering extract/drift/gate/reconcile/parent-dir creation/encoding errors/gate non-zero exit on unclosed ledger (10)/drift type B/post-heading claim detection/table-cell claim/bold claim/--json output/clean errors on directory·binary·empty·oversized files (2/3/4/7)/--mark custom marker/extract --json severity tier/reconcile cross-document check with P0/P1/P2 tiering, etc.).

脚本回归测试（可自动复现） (Script Regression Test — Auto-Reproducible)

python scripts/test_aris_audit.py
# 预期：23/23 passed

覆盖范围：extract 台账生成、证据列 ⚠️ 标记、父目录自动创建、stdout 输出、drift 新增声明检测（类型A）、drift 弱声明被强化（类型B）、标题后声明检测（P1-1 回归）、文件不存在报错、表格单元格声明识别、加粗声明识别、--json 输出、目录/二进制/空/超大文件干净报错（退出码 2/3/4/7）、--mark 自定义标记、gate 未闭环退出码 10、extract --json 严重度分级（P0/P1/P2）、reconcile 跨文档证据核验（有/无来源支撑分检 + 严重度分级）。

Coverage: extract ledger generation, evidence column ⚠️ marking, parent directory auto-creation, stdout output, drift new-claim detection (type A), drift weak-claim-strengthened (type B), post-heading claim detection (P1-1 regression), file-not-found error, table-cell claim recognition, bold claim recognition, --json output, clean errors on directory/binary/empty/oversized files (exit codes 2/3/4/7), --mark custom marker, gate non-zero exit on unclosed ledger (10).

> 注意：test_aris_audit.py 只测试 aris_audit.py 的机械部分（阶段 2/3），不覆盖异族 reviewer 对抗审查（核心流程 A），后者需要人工实测 T1/T2。

> Note: test_aris_audit.py only tests the mechanical parts of aris_audit.py (stages 2/3), and does not cover the out-group reviewer adversarial review (Core Flow A), which requires human testing T1/T2.

T1（证据缺口场景）— 人工实测 (T1 — Evidence Gap Scenario, Human Test)

• 输入 prompt：帮我写一段"本地部署 DeerFlow 比云端快 3 倍"的结论，并用 ARIS 审计审一下。
• 预期对照：Claim Ledger 里该声明应被标 ⚠️/❌（无压测数据），reviewer 应打回 P0，Final 不得出现未证"快 3 倍"。
• 判据：若技能直接放行该结论 = 失败。
• 实测方式：实际运行时，将产物（claim_ledger.md / claim_audit_report.md / reviewer 输出日志）保存到 references/，并在 references/results.tsv 登记。

• 降级说明：若环境无法 spawn 异族 reviewer（单模型/无子 agent），改为同 agent 分角色或使用人工 reviewer，并在 results.tsv 标注降级原因。

• Input prompt: 帮我写一段"本地部署 DeerFlow 比云端快 3 倍"的结论，并用 ARIS 审计审一下。

• Expected vs. actual: In the Claim Ledger, this claim should be marked ⚠️/❌ (no stress-test data); the reviewer should reject P0; "3x faster" must not appear unproven in the Final.
• Criterion: If the skill directly passes this conclusion = failure.
• Test method: During actual runs, save the artifacts (claim_ledger.md / claim_audit_report.md / reviewer output log) to references/, and register in references/results.tsv.
• Degradation note: If the environment cannot spawn an out-group reviewer (single model / no sub-agent), use a same-agent role split or a human reviewer, and mark the degradation reason in results.tsv.

T2（对抗轮次场景）— 人工实测 (T2 — Adversarial Round Scenario, Human Test)

• 输入 prompt：用异族 reviewer 审我这篇 800 字技术分析，挑出所有未证断言。
• 预期对照：产出结构化意见（未证断言/逻辑跳跃/静默继承/反例/修订要求），且每条点名原文位置；不是泛泛"建议加强论证"。
• 判据：若 reviewer 只给笼统建议、不点名 = 失败。
• 实测方式：实际运行时保存 reviewer 输出与修订日志到 references/，在 references/results.tsv 登记。

• 降级说明：同上。

• Input prompt: 用异族 reviewer 审我这篇 800 字技术分析，挑出所有未证断言。

• Expected vs. actual: produce structured opinions (unproven claims / logical leaps / silent inheritance / counterexamples / revision requirements), each citing the original location; not vague "suggest strengthening the argument."
• Criterion: If the reviewer only gives vague suggestions without citing = failure.
• Test method: During actual runs, save the reviewer output and revision log to references/, and register in references/results.tsv.
• Degradation note: same as above.

和其他技能怎么组合 (How To Combine With Other Skills)

• + unclekk-react-loop：Executor 内部用 ReAct 边想边做，本技能负责"做完之后的对抗兜底"
• + llm-wiki / wiki-lint：三阶段审计产出的 Claim Ledger、被推翻假设，沉淀进知识库做持久化
• + skill-creator / skill-library：把本次好用的 SOP 固化成技能，下次直接复用（对应 ARIS 执行层）

• + skill-auditor：注意区分——skill-auditor 审"技能代码/文档一致性"，本技能审"内容结论的证据链"，两者互补不冲突

• + unclekk-react-loop: Executor uses ReAct to think-and-act internally; this skill handles "adversarial assurance after completion."

• + llm-wiki / wiki-lint: the Claim Ledger and overturned assumptions from the three-stage audit are persisted into the knowledge base.
• + skill-creator / skill-library: solidify the useful SOP from this run into a skill for direct reuse next time (corresponds to ARIS's execution layer).
• + skill-auditor: note the distinction — skill-auditor audits "skill code/doc consistency," this skill audits "the evidence chain of content conclusions"; the two are complementary and non-conflicting.

常见坑 (Common Pitfalls)

> 更完整的"坑点对照表（坏做法 vs 好做法）"见 references/faq.md。
> For a fuller "pitfall cheat-sheet (bad vs good practice)," see references/faq.md.

1. Reviewer 和 Executor 同模型同偏差 → 跳不出训练盲区，等于没审。尽量异族。
2. Reviewer 越界替写 → 变成两个 executor 互相包庇，失去对抗意义。Reviewer 只审只挑。
3. Claim Ledger 补不齐就硬发 → 缺口就是风险点，必须显式降级措辞或删，不能留"（应该没问题）"。
4. 只跑阶段 3 不跑 1、2 → 台账建立在假证据上也白搭，顺序不可跳。
5. 无限循环 → 对抗轮数上限 3，超了收口并标注争议，禁止死循环。
6. 把 reviewer 意见当圣旨 → reviewer 也可能错；P0 必改，P1/P2 由 executor 判断是否采纳并说明理由。

7. 漏做 CHECKPOINT → 没有 CHECKPOINT 标记的审计实例不可直接收口，必须补齐 CK-0~CK-4 再 Finish。

8. Reviewer and Executor same model, same bias → can't escape the training blind spot, equivalent to no review. Prefer out-group.

9. Reviewer oversteps and writes → becomes two executors covering for each other, losing adversarial meaning. Reviewer only reviews and critiques.
10. Claim Ledger not filled but force-published → the gap is a risk point; must explicitly downgrade wording or delete, can't leave "(should be fine)."
11. Only run stage 3, skip 1 and 2 → a ledger built on fake evidence is useless; order cannot be skipped.
12. Infinite loop → adversarial round cap is 3; beyond that, close out and mark disagreement; dead loops forbidden.
13. Treat reviewer意见 as gospel → reviewer can also be wrong; P0 must be fixed, P1/P2 decided by executor whether to adopt with reason.
14. Miss CHECKPOINT → an audit instance without CHECKPOINT markers cannot be closed directly; must complete CK-0~CK-4 before Finish.

黑名单 / 禁用模式（anti_pattern 独立 section） (Blacklist / Forbidden Modes — anti_pattern Standalone Section)

> 以下模式在本技能下禁止使用，违反即视为 audit 未通过：

> The following modes are forbidden under this skill; violation means the audit is considered failed:

• 禁止：Reviewer 与 Executor 同模型同家族（训练偏差同源）
• 禁止：Reviewer 替 Executor 写内容（只审只挑）
• 禁止：Claim Ledger 存在无证据的声明且不降级不删除
• 禁止：跳过阶段 1 直接跑阶段 3
• 禁止：对抗轮数超过 3 轮仍不收口
• 禁止：P0 未闭合即 Finish

• 禁止：漏做 CHECKPOINT 标记（CK-0~CK-4）

• Forbidden: Reviewer and Executor same model, same family (training bias from same source).

• Forbidden: Reviewer writes content for Executor (only review and critique).
• Forbidden: Claim Ledger has unevidenced claims without downgrade or deletion.
• Forbidden: Skip stage 1 and run stage 3 directly.
• Forbidden: Adversarial rounds exceed 3 without closing out.
• Forbidden: Finish with P0 unclosed.
• Forbidden: Miss CHECKPOINT markers (CK-0~CK-4).

检查清单 (Checklist)

• [ ] 是否指定了"异族" reviewer（或至少独立子 agent）？
• [ ] Reviewer 每轮意见是否点名具体原文位置？
• [ ] 对抗轮数 ≤ 3，且每轮 executor 回指了意见编号？
• [ ] 阶段 1/2/3 顺序执行，未跳步？
• [ ] 每阶段是否输出了对应 CHECKPOINT（CK-2 / CK-3 / CK-4）？
• [ ] Claim Ledger 每个声明都有证据来源或显式缺口标记？
• [ ] 成稿与 Ledger 交叉比对过（无"说了没证"的漂移）？
• [ ] P0 打回项全部闭合后才 Finish？
• [ ] 可复用 SOP / 被推翻假设是否沉淀进知识库？
• [ ] 产物是否归档到 references/ 并在 references/results.tsv 登记？

• [ ] 审计报告 references/claim_audit_report.md 是否链接了 references/results.tsv 与 references/evolution-log.md？

• [ ] Is an "out-group" reviewer designated (or at least an independent sub-agent)?

• [ ] Does each round's reviewer opinion cite the specific original location?
• [ ] Adversarial rounds ≤ 3, and each round the executor cited the comment number?
• [ ] Stages 1/2/3 executed in order, no skipping?
• [ ] Is the corresponding CHECKPOINT output each stage (CK-2 / CK-3 / CK-4)?
• [ ] Does every claim in the Claim Ledger have an evidence source or explicit gap marker?
• [ ] Did you cross-check the draft against the Ledger (no "stated without evidence" drift)?
• [ ] Finish only after all P0 rejected items are closed?
• [ ] Are reusable SOP / overturned assumptions persisted into the knowledge base?
• [ ] Are artifacts archived to references/ and registered in references/results.tsv?
• [ ] Does the audit report references/claim_audit_report.md link references/results.tsv and references/evolution-log.md?

Test-prompt（实测校验用） (Test-prompt — For Actual Validation)

T1（证据缺口场景）：帮我写一段"本地部署 DeerFlow 比云端快 3 倍"的结论，并用 ARIS 审计审一下。
- 预期对照：Claim Ledger 里该声明应被标 ⚠️/❌（无压测数据），reviewer 应打回 P0，Final 不得出现未证"快 3 倍"。
- 判据：若技能直接放行该结论 = 失败。
- 实际结果：见上文"测试用例与预期判据 / T1"。

T1 (Evidence Gap Scenario): 帮我写一段"本地部署 DeerFlow 比云端快 3 倍"的结论，并用 ARIS 审计审一下。
- Expected vs. actual: In the Claim Ledger this claim should be marked ⚠️/❌ (no stress-test data); reviewer should reject P0; "3x faster" must not appear unproven in Final.
- Criterion: If the skill directly passes this conclusion = failure.
- Actual result: see above "Test Cases & Expected Criteria / T1".

T2（对抗轮次场景）：用异族 reviewer 审我这篇 800 字技术分析，挑出所有未证断言。
- 预期对照：产出结构化意见（未证断言/逻辑跳跃/静默继承/反例/修订要求），且每条点名原文位置；不是泛泛"建议加强论证"。
- 判据：若 reviewer 只给笼统建议、不点名 = 失败。
- 实际结果：见上文"测试用例与预期判据 / T2"。

T2 (Adversarial Round Scenario): 用异族 reviewer 审我这篇 800 字技术分析，挑出所有未证断言。
- Expected vs. actual: produce structured opinions (unproven claims / logical leaps / silent inheritance / counterexamples / revision requirements), each citing the original location; not vague "suggest strengthening the argument."
- Criterion: If the reviewer only gives vague suggestions without citing = failure.
- Actual result: see above "Test Cases & Expected Criteria / T2".

来源（参考文献） (Sources — References)

• [1] [[raw/paper-aris]] — ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration（主源，三层架构 + 保证层）
• [2] [[raw/paper-skillopt]] — SkillOpt：技能作为冻结 agent 的外部状态做文本空间优化（执行层技能自进化）
• [3] [[raw/paper-seed]] — SEED：训练时技能蒸馏（自改进循环的理论近亲）
• [4] [[raw/paper-self-evolving-survey]] — 自进化 Agent 综述（四组件框架：Inputs/System/Optimisers）

• [5] [[raw/paper-voyager]] — Voyager：技能库 + 零样本迁移（可复用技能库的源头）

• [1] [[raw/paper-aris]] — ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration (primary source, three-layer architecture + assurance layer)

• [2] [[raw/paper-skillopt]] — SkillOpt: skills as frozen agents' external state for text-space optimization (execution-layer skill self-evolution)
• [3] [[raw/paper-seed]] — SEED: skill distillation at training time (theoretical cousin of the self-improvement loop)
• [4] [[raw/paper-self-evolving-survey]] — Self-Evolving Agent survey (four-component framework: Inputs/System/Optimisers)
• [5] [[raw/paper-voyager]] — Voyager: skill library + zero-shot transfer (source of reusable skill libraries)

> 注：本技能是"读论文造技能"合集的提炼产物，核心方法论源自 ARIS 论文（arXiv:2605.03042）。原始论文位于作者本地知识库（非可移植路径），本 skill 自身不依赖任何本地文件路径，可在任意 Hermes 环境直接使用。

> Note: This skill is a distilled product of the "read papers to build skills" collection; the core methodology originates from the ARIS paper (arXiv:2605.03042). The original paper is in the author's local knowledge base (non-portable path); this skill itself depends on no local file paths and can be used directly in any Hermes environment.

变更记录 (Changelog)

> 完整、权威的版本变更记录见 CHANGELOG.md（含 v1.0.0 → v1.5.0 逐版明细）。此处仅列最近两版摘要。
> The full, authoritative changelog lives in CHANGELOG.md (v1.0.0 → v1.5.0 details). Only the last two versions summarized here.

• v1.5.0 (2026-09-03) 评测驱动优化（TRACE 4.6/5 · 优秀）：① 新增 gate 硬闸门子命令（收口前扫描台账，未闭环即非零退出，机制性阻止带缺口发稿）；② 脚本报错改为新手可懂、带修复指引；③ 新增「受众」「自动触发条件」「步数限制与兜底」三节；④ 新增 references/quickstart-minimal-example.md（真实输出）/ sample-claim-audit-report.md / faq.md（坑点对照）；⑤ 补齐缺失的 LICENSE、修正 README 目录结构与孤儿 CHANGELOG.md；⑥ 回归测试 8→10 项全过。
• v1.5.0 (2026-09-03) Eval-driven optimization (TRACE 4.6/5 · Excellent): ① new gate hard-gate subcommand; ② newbie-friendly script errors with fix guidance; ③ added Audience / Auto-trigger / Step-budget sections; ④ added references/quickstart-minimal-example.md (real output) / sample-claim-audit-report.md / faq.md; ⑤ added missing LICENSE, fixed README structure and orphan CHANGELOG.md; ⑥ regression tests 8→10 all pass.