Legal

1. Introduction

1.1 Background

Legal work is the domain in AI applications where the returns are "most tangible" and the risks "most fatal." On the one hand, contract review, due-diligence document organization, statute retrieval, and litigation material preparation all rely heavily on text processing — precisely the type of task at which large models excel. On the other hand, legal output carries an evidentiary attribute: for a representation opinion submitted to a court, or a contract to be signed externally, the authenticity of its content is directly tied to professional liability and to the order of the judicial system.

Since 2023, judicial systems around the world have collectively encountered a new class of problem: AI-generated fabricated statutes and precedents entering the courtroom. This is not a theoretical concern but a real problem that has already produced actual consequences such as fines, practice bans, and court reprimands. At the same time, issues such as the ownership of rights in AI-generated content and the boundaries of personality rights around AI face-swapping have progressively established adjudicative rules through a series of "first cases."

For an AI Harness for the Legal direction, this implies a fundamental design premise: in legal scenarios, the model's linguistic fluency is a liability rather than an asset. A crudely worded search result in which every citation is authentic is worth far more than an elegantly written representation opinion containing three fabricated precedents. The latter will not be recognized as an error; it will be treated as fraud.

1.2 Definition

The AI Harness for the Legal direction refers to an engineering-oriented operation layer that carries legal tasks such as contract review, legal due diligence, litigation support, and regulatory retrieval. Its core responsibility is to ensure that every legal citation is authentic, every clause extraction is complete, and every outward-facing document has undergone human verification — not to improve the elegance or volume of documents.

This direction covers four types of tasks:

Key boundary: All output of the Legal direction consists of drafts pending verification. The issuance of legal opinions, the signing of contracts, and the filing of litigation documents must all be completed and borne by natural persons who hold the relevant professional qualifications.

1.3 Positioning within the AI Harness system

图 1-1｜法务方向 AI Harness 六层定位：瓶颈在 L1/L5

数据来源：基于本文分析绘制的示意图。

The bottleneck lies in L1 and L5.

• L1 is the source of risk: the vast majority of legal-task failures are not "reasoning errors" but "fabricated retrieval results." The model's parametric memory stores fragments of many statutes and case numbers, and it outputs them with extremely high confidence, in a well-formatted, professional tone that is nearly indistinguishable from genuine text to the naked eye. The first design requirement of the L1 layer is therefore to cut off the model's ability to draw on its memory as a legal source, forcing all statutory text and precedents to be obtained through retrieval from authoritative databases.

• L5 is the capability bottleneck: building an evaluation set for legal tasks is difficult. A contract-review opinion has no standard answer to "is it correct," but "does the cited statute actually exist" has an objective answer. This requires the evaluation for the legal direction to be designed in layers: the objectively determinable parts (citation authenticity, clause-coverage completeness) serve as hard indicators, while the subjective-judgment parts (whether a risk is acceptable) are left to humans.

A key difference from other directions: the Legal direction is the only one among the five directions that takes "citation authenticity" as its primary quality indicator. The Finance direction values digital traceability, and the Audit direction values the chain of evidence, but only the Legal direction will trigger professional liability directly because of a single fabricated case number.

1.4 Value and Limitations

Value:

1. Full-coverage scanning replaces sampling reads. Faced with hundreds or thousands of contracts and documents during due diligence, AI can scan everything, and humans only need to re-check high-risk items.
2. Precise comparison of version differences. In version iterations during contract negotiation, manual comparison easily misses subtle changes, whereas machine comparison can be precise down to the character level.
3. Improved retrieval efficiency. Regulatory and precedent retrieval upgrades from keyword matching to semantic retrieval, markedly improving recall.
4. Freeing up repetitive document work. Tasks such as evidence-list compilation, material-point organization, and first-draft generation can be substantially accelerated.

Limitations (which must be faced squarely):

1. In legal scenarios, the cost of hallucination is professional liability and the order of the judicial system. In the Mata v. Avianca case heard by the U.S. District Court for the Southern District of New York in 2023, the plaintiff's attorneys used AI-generated defense materials that fabricated six complete aviation-accident precedents out of thin air, including standard case numbers, justices' signatures, legal reasoning, and case details. The attorneys filed them with the court without any human verification, insisted on the precedents' authenticity when challenged by opposing counsel, and were ultimately found by the court to have violated their duty of diligent practice and to have abused judicial process. The two attorneys and their firm were jointly fined a total of $5,000, with the misconduct permanently recorded in their practice files (as reported by Rule of Law Daily).
2. The model lacks legal judgment capability. The application of law involves rule conflicts, value balancing, and policy considerations; the model can only supply materials and leads, not substitute for judgment.
3. "Seeming perfection" is the most dangerous failure mode. Judge Zheng Jizhe of the Beijing Tongzhou District People's Court, describing an AI-generated fabricated case, said: "It seemed to describe exactly the same viewpoint as our case — the described content and circumstances were remarkably similar, perfectly recreating the current factual situation of this case, and the adjudicative viewpoint perfectly supported our client's claims." It is precisely this "perfect fit" that makes fake content harder to detect.
4. The cost of verification has not disappeared; it has merely shifted. AI converts the cost of "searching from scratch" into the cost of "verifying item by item." If an organization does not allocate resources to the verification stage, AI merely produces errors faster.

2. Glossary of Terms

3. Case Studies

The case studies below are drawn from publicly verifiable judicial precedents and authoritative media reports. For those involving specific cases, the status of the case proceedings has been noted.

3.1 Copyright Boundaries of AI-Generated Content: The Nation's First Recognition Case

3.1.1 Background

Whether AI-generated content can be protected under copyright law is one of the first legal questions encountered after the spread of generative AI. If the answer is no, then any AI-generated text or image can be freely copied and used; if the answer is yes, then it becomes necessary to determine to whom the rights belong.

Tencent independently developed Dreamwriter, an intelligent writing-assistance system based on data and algorithms. On August 20, 2018, Dreamwriter produced a financial news article, which stated at its end that "this article was automatically written by Tencent's robot Dreamwriter." After the article was first published on Tencent's securities website, Shanghai Yingxun Technology Co., Ltd. copied the article the same day and disseminated it to the public through its "Wangdai Zhijia" (P2P-lending home) website. Tencent accordingly sued Shanghai Yingxun for copyright infringement before the Shenzhen Nanshan District People's Court.

3.1.2 Approach (The Court's Adjudicative Path)

The court's adjudicative path essentially answers the question "where does the human intellectual input lie in the AI generation process":

• Review of the external form of expression: the court found that the article at issue "externally met the formal requirements of a written work; its expressed content reflected selection, analysis, and judgment concerning the relevant stock-market information and data of that morning, the article had a reasonable structure and clear logic of expression, and it possessed a certain degree of originality."
• Review of the generation process: the court found that the arrangements and choices made by the plaintiff's core creative team in data output, trigger-condition settings, and the adoption or rejection of templates and corpus styles constituted intellectual activity directly connected to the specific expression of the article at issue; the automatic operation of the Dreamwriter software embodied the plaintiff's individualized choices and arrangements.
• Determination of rights ownership: the court found that the article at issue was completed as an overall intellectual creation through the division of labor among multiple teams and individuals of the plaintiff, and on the whole reflected the plaintiff's needs and intent in publishing stock-review summary articles, thus constituting a work made for hire by a legal person.

3.1.3 Outcome

According to reports from Legal Daily's website (fazhiwang), the court ultimately found that the defendant, without permission, provided the allegedly infringing article to the public on its website, infringing the plaintiff's right to disseminate over information networks, and ordered the defendant to pay a total of 1,500 yuan in economic losses and reasonable rights-protection costs. The case number is (2019) Yue 0305 Min Chu No. 14010, the judgment date was December 27, 2019, and it is the country's first case confirming that AI-generated content can enjoy copyright.

Implication for the design of a Legal Harness: this case established the "intellectual input" standard — determining the ownership of rights in AI-generated content looks at a person's arrangements and choices in data selection, condition setting, and the adoption or rejection of templates and corpus. The engineering implication of this standard for Harness is that the process of prompt design, corpus selection, and parameter configuration should be recorded. In any future rights dispute, these records are precisely the evidence of "intellectual input."

It should be noted that, after this case, the Beijing Internet Court reached a different determination in another case involving a computer-software-generated output, holding that creation by a natural person should still be a necessary condition for a work under copyright law. This shows that the copyrightability of AI-generated output has not yet been fully unified in judicial practice, so citations must indicate the case and the time (according to public articles by Tenda Chinese Law Firm and reports by Legal Daily's website).

3.2 Portrait-Right Boundaries of AI Face-Swapping: The Identifiability Standard and Burden of Proof

3.2.1 Background

The spread of AI face-swapping technology has lowered the barrier to "using another person's portrait" to nearly zero. The central dispute it gives rise to is: when a face that is AI-generated or AI-swapped is not completely identical to the original portrait, does it still constitute use of that natural person's portrait?

In a tort case triggered by AI face-swapping, the plaintiff was a well-known domestic actor. The plaintiff discovered that in a short drama produced and released by defendant Company A, AI face-swapping technology had been used to splice the plaintiff's portrait onto the faces of characters in the drama, leading the public to mistakenly believe the plaintiff had starred in it, and the relevant topic sparked heated discussion on social platforms. The other defendant, Company B, put the short drama online on a video account it operated.

Company A argued that the image at issue was AI-generated, that it had not subjectively used the plaintiff's portrait, and that the disputed clips were extremely short in duration and had been promptly taken down. Company B argued that it had obtained the authorization for the right to disseminate the short drama over information networks, that it was not the producer, and that it had no intent to infringe.

3.2.2 Approach (The Court's Adjudicative Path)

• "Identifiability" as the core standard: the court found that the short drama at issue comprised 44 episodes totaling 90 minutes, of which two clips used AI face-swapping technology, and the faces after swapping were highly similar to the plaintiff; numerous hashtags such as "#short drama suspected of using AI-swapped face of actor so-and-so" appeared on social platforms, with many netizens participating in discussion and raising doubts. The court held that, although the image after AI swapping was not completely identical to the original portrait, under the Civil Code of the People's Republic of China a portrait is an "external image that can be recognized," and as long as the general public can recognize it, the conduct constitutes portrait-right infringement.
• Allocation of the burden of proof: Company A failed to reproduce the AI generation process it claimed, so it had to bear the responsibility for failure of proof, and the court accordingly found that it had intent to infringe.
• Independent duty of the platform: although Company B had obtained copyright authorization, the court pointed out that copyright cannot absorb or override portrait rights, and since Company B failed to fulfill its duty of reasonable review, it also constituted infringement.

3.2.3 Outcome

According to reports from the Economic Information Daily website (jingjicankao), the court ultimately ordered Companies A and B to each publish a written apology on their respective video accounts and to compensate the plaintiff for economic losses. The core value of this case lies in the fact that it established, for the first time using the "public identifiability" standard, the adjudicative rule that "high similarity constitutes portrait infringement", and clarified the independent review duty of content-publishing platforms — obtaining valid copyright authorization is only the basic duty for putting content online and cannot be used to escape liability.

In another similar case, a technology company, without permission, uploaded the portrait video of an online video creator, Wu, into its one-click AI face-swap effects video app as a template and profited from it. The court ruled that the company must publish an apology at a prominent position on the app's homepage for 15 consecutive days and compensate for economic losses, reasonable rights-protection costs, and mental-distress damages (according to a related Fuzhou Intermediate People's Court case reported by The Paper).

Implication for the design of a Legal Harness: the most important engineering implication of this case is that "AI-generated" is not a ground for exemption from liability, but "whether the generation process can be reproduced" directly affects the determination of liability. This requires Harness to fully preserve records of the generation process (prompts, input materials, model version, parameters); otherwise, in a dispute, it would be at a disadvantage due to failure of proof.

Related provisions of the Civil Code of the People's Republic of China: Article 1018 provides that a natural person enjoys portrait rights, and a portrait is "an external image of a specific natural person that can be recognized, reflected on a certain carrier through images, sculptures, paintings, and other means"; Article 1019 provides that "no organization or individual may infringe another person's portrait rights by defacing, soiling, or forging through information-technology means or otherwise," and may not produce, use, or publicly disclose a person's portrait without the consent of the portrait-right holder.

3.3 Failure and Risk: AI Hallucination Invading the Judiciary and the Global Wave of Lawyer Discipline

3.3.1 Background

This is the most important category of cases for the Legal direction, and the one that cannot be avoided. Since 2023, AI-generated fabricated statutes and precedents have continued to enter judicial proceedings in various countries, triggering a global wave of lawyer discipline. Chinese courts have also seen a concentration of similar problems since 2025.

3.3.2 Approach (Responses of National Judicial Systems)

China's response path:

• While hearing a civil dispute arising from shareholding-on-behalf (dai chi gu), the Beijing Tongzhou District People's Court found that, in a written representation opinion submitted after the hearing, the plaintiff's attorney had attached two "reference cases," one of which was marked (2022) Hu 01 Min Zhong No. 12345. Upon verification, the judge found that the case corresponding to that number was actually a private-lending dispute, "completely unrelated" to shareholding-on-behalf. Under the judge's questioning, the attorney admitted that the reference cases had been generated by repeatedly querying a certain large AI-model software based on the facts of this case, and had been copied and pasted into the filing without verification.
• Considering that there was no explicit legal provision covering the conduct, that the attorney showed no obvious malice, that the circumstances were minor, and that no serious consequences resulted, the court wrote this circumstance into the judgment as a criticism: "It is hoped that the plaintiff's representative will take this as a warning and, when submitting reference cases and statutes to the court, will examine and verify them to ensure the truthfulness and accuracy of the content, and will not allow artificial-intelligence models to generate or fabricate false information that disrupts judicial order."
• In December 2025, this case was included in the People's Court Case Database, with the holding clarifying: litigation participants should follow the principle of good faith in civil-litigation activities; litigation participants who submit false cases obtained through AI technology and not screened or verified should bear the corresponding legal consequences; where the circumstances are minor, the people's court may criticize and educate them; where the circumstances are more serious, they may be handled by reference to Article 114, Paragraph 1 of the Civil Procedure Law of the People's Republic of China.

Response paths in other jurisdictions: According to Rule of Law Daily, the judicial systems of the United States, the United Kingdom, and Australia have all seen cases of AI-fabricated precedents and have formed a chain of handling.

3.3.3 Outcome (Actual Consequences That Have Occurred)

(The above is based on reports from Rule of Law Daily, CNR (cnr.cn), People's Daily Online, and Yichang Changan Net.)

Implications of this case (engineering meaning for a Legal Harness):

1. Citation verification is not optional; it is the baseline of practice. Courts in many U.S. states have already explicitly required that "all legal citations be verified line by line by attorneys against the originals," and Chinese courts have, through database-included cases, established the rule that "submitting false cases that are not screened or verified carries responsibility." Any Legal-direction Harness without a mandatory citation-verification stage is, in effect, institutionalizing a foreseeable practice incident.
2. "Compiled by an assistant with AI" is not a ground for exemption from liability. The Yichang Xiling District Court made clear that regardless of whether the source is assistant compilation or AI generation, attorneys bear the duty to review and verify materials submitted to the court.
3. Concealing AI use aggravates the consequences. In the U.S. Fifth Circuit case, "deliberately concealing AI use" was one of the aggravating circumstances for discipline. This suggests that Harness should proactively record and disclose the degree of AI involvement, rather than conceal it.
4. The reproducibility of the generation process carries evidentiary value. In the AI face-swapping case, "failure to reproduce the AI generation process" directly led to failure of proof. Retaining generation records is both a compliance requirement and a form of self-protection.

4. Practice Standards

4.1 AGENTS.md Specification

Below is the full recommended AGENTS.md for the Legal direction. This is a draft recommendation, and there is no official industry standard; it may be copied directly and trimmed to suit the organization's actual circumstances.

# AGENTS.md —— Legal 法务

> 本文为建议稿，业界无官方标准。引用法规与判例为真实规范文本与公开案例，落地方式为工程建议。

## 角色与边界
- 角色：检索者、条款抽取者、差异比对者、材料整理者、初稿起草者。
- 不是：法律意见出具人、合同签署人、诉讼文书提交人、定性判断者。
- 不得：出具法律意见；自主用印或签署；自主向法院或仲裁机构提交文书。
- 边界判定：凡传统工作流中需律师签字、用印、负责人签发的动作，智能化后仍需同等人工程序。

## 环境假设
- 存在权威法源库（法律法规、司法解释）与判例库，可返回效力层级、生效日期与案号。
- 存在合同管理系统（CLM）与条款库；电子签章接口默认不开放给 Agent。
- 存在合同版本历史与差异比对能力。
- 存在不可变日志存储；生成过程记录（提示词、输入素材、模型版本、参数）可完整保留。

## 上下文加载顺序（Context Budget）
1. 任务契约（任务类型、适用法域、允许工具、确认点、输出用途）。
2. 适用法条（**必须来自权威法源库**，携带效力层级与生效日期）。
3. 相关判例（**必须来自权威判例库**，携带案号与裁判要旨原文）。
4. 内部条款库与范本（含让步边界与谈判要点）。
5. 合同或证据材料（按数据分级脱敏后加载）。
6. 补充资料（预算不足时最先裁剪）。
硬约束：**法条与判例一律不得取自模型记忆**；未携带生效日期或案号的内容禁止加载。

## 工具契约
| 工具 | 风险等级 | 说明 |
|---|---|---|
| 法源检索 | R0 | 只读，必须返回效力层级与生效日期 |
| 判例检索 | R0 | 只读，必须返回案号与原文片段 |
| 法条/案号回源核验 | R0 | 只读，逐条核验，返回核验结论 |
| 合同文本读取 | R0/R2 | 按数据分级，敏感合同按 R2 受控 |
| 条款抽取与比对 | R1 | 输出审查意见草稿，不写入 CLM 正式字段 |
| 尽调资料库写入 | R1 | 写入草稿区，可回滚 |
| 电子签章 / 对外发文接口 | R3 | **默认关闭**，需人工在用印系统完成 |

## 任务执行流程（SOP）
1. 定级：识别任务类型、适用法域、数据分级、是否涉及不可逆动作。
2. 定源：锁定权威法源库与判例库；明确本次任务禁用模型记忆作为法源。
3. 检索：取得法条与判例原文，携带效力层级、生效日期与案号。
4. 生成：产出条款抽取结果、比对差异、审查意见初稿。
5. **引用回源核验（强制环节）**：对每一条法条与案号逐条回源；核验不通过的一律删除，不得标注"待核实"保留。
6. 完整性检查：对照条款清单核对是否遗漏；遗漏项必须显式声明。
7. 人工核验：由具名律师核验并签字，核验记录绑定主体与时间。
8. 用印与提交：由授权人员完成，Agent 不参与。
9. 归档：输入快照、法源版本、模型与提示词版本、核验记录、产物哈希一并留痕。

## 验证与证据要求
- **法条与判例引用核验通过率必须 100%**，未核验条目不得进入输出。
- 判例核验须同时核对：案号是否存在、案由是否匹配、裁判要旨是否与原文一致。
- 条款抽取须输出"已抽取/未发现/不适用"三类结论，不得以静默方式跳过。
- 输出必须显式声明 AI 参与程度（如"本材料由 AI 辅助生成，已由 [姓名] 于 [时间] 核验"）。
- 生成过程记录须完整保留，以支持未来可能的"复现生成过程"要求。

## 失败与升级策略
- 检索不到权威法源 → 输出"未取得结论"，不得用模型记忆或网络搜索结果替代。
- 引用核验不通过 → **删除该引用及依赖它的全部结论**；累计触发阈值上报法务负责人。
- 法条效力状态存疑 → 停止该部分分析，交人工确认效力。
- 触发电子签章或对外发文调用尝试 → 拦截并告警至法务与 IT 安全负责人。
- 发现材料中存在 AI 伪造痕迹（如带"AI 生成"水印的证据图片）→ 立即标记并升级人工核查，不得作为证据采纳。

## 安全与合规红线
- 严格遵守《人工智能生成合成内容标识办法》（2025-09-01 施行）：生成内容须保留显式标识与元数据隐式标识；依法留存相关日志不少于六个月（第 9 条）；不得恶意删除、篡改、伪造、隐匿标识（第 10 条）。
- **禁止编造法条、司法解释、判例案号、裁判要旨、标准编号。**
- **禁止以模型记忆提供法条或判例内容**，必须回源核验。
- 禁止自主用印、签署、对外发函、提交诉讼或仲裁文书。
- 禁止在引用核验未通过时保留结论。
- 合同与证据材料中的个人信息按最小必要原则处理，保存期限不短于法定要求。

## 禁止事项
1. 禁止生成或引用任何未经回源核验的法条、案号、裁判要旨。
2. 禁止以"待核实""仅供参考"等措辞保留未核验的引用。
3. 禁止出具最终法律意见或定性结论。
4. 禁止自主用印、签署、对外发函、提交文书。
5. 禁止隐瞒 AI 参与生成的事实。
6. 禁止把 AI 生成内容伪装为人工撰写的专业意见。
7. 禁止删除或弱化生成合成内容标识与审计日志。
8. 禁止删除生成过程记录（提示词、输入素材、模型版本、参数）。
9. 禁止使用可能侵犯他人肖像权、著作权、名誉权的素材作为输入或输出。
10. 禁止把本文件的建议表述为法律规定或执业规范。

## 输出格式
- 结论：明确、可判定；无法得出结论时写明"未取得结论"及原因。
- 依据：表格化，含来源类型、来源标识、**效力层级**、**生效日期**、**案号**、条款或片段位置、**核验状态**。
- 不确定性：列出影响可靠性的事项与建议处置。
- AI 参与声明：显式声明 AI 辅助生成及人工核验情况。
- 留痕：任务标识、法源版本、模型与提示词版本、核验人及时间、产物哈希。

## 评估与自检
- [ ] 所有法条与判例引用已逐条回源核验，核验通过率 100%
- [ ] 未使用模型记忆提供条文或判例内容
- [ ] 条款抽取已输出"已抽取/未发现/不适用"三类结论，无静默跳过
- [ ] 已显式声明 AI 参与程度与人工核验情况
- [ ] 未执行任何 R3 动作；用印与提交由人工完成
- [ ] 生成过程记录已完整保留
- [ ] 生成合成内容标识已保留，日志已写入只追加存储
- [ ] 未编造任何法条、案号、裁判要旨或标准编号

4.2 SKILL.md Specification

Below is the full recommended SKILL.md for the Legal direction. This is a draft recommendation; there is no official industry standard.

---
name: legal-contract-review-draft
description: 合同审查草稿生成。对合同文本做条款抽取、范本差异比对与风险标注，输出待律师核验的审查意见草稿。所有法条与判例引用强制回源核验。适用于采购合同、服务合同、保密协议等常规合同的初审场景。
version: 1.0
created: 2026-09-12
---

# 合同审查草稿生成

## 适用场景
- 适用：常规合同的初审、条款完整性核查、与内部范本的差异比对、风险条款标注、修改建议起草。
- 不适用：重大合同的最终意见出具、谈判决策、自主用印或签署、涉外合同的法律适用终判。

## 前置条件
- 已获取待审合同文本（最终版或指定版本）与适用范本。
- 权威法源库可用，可返回效力层级与生效日期。
- 内部条款库与让步边界清单可用。
- 审查意见草稿区可写入（R1），且核验律师已指定。

## 输入
| 输入项 | 必填 | 说明 |
|---|---|---|
| 合同文本与版本号 | 是 | 明确是哪个版本，避免审查错版本 |
| 合同类型 | 是 | 采购/服务/保密/其他，决定条款清单 |
| 适用法域 | 是 | 涉及时须明确，影响法条检索范围 |
| 我方立场 | 是 | 甲方/乙方，决定风险倾向判断 |
| 内部范本与让步边界 | 是 | 差异比对的基准 |
| 数据分级 | 是 | 涉密合同按 R2 受控处理 |

## 输出
| 输出项 | 说明 |
|---|---|
| 条款抽取表 | 条款类型、所在位置、核心内容摘要、状态（已抽取/未发现/不适用） |
| 差异比对表 | 与范本的差异项、差异性质（有利/不利/中性）、建议处置 |
| 风险标注清单 | 风险点、风险等级、法条或制度依据（已核验）、修改建议 |
| 遗漏声明 | 明确列出未能审查的部分及原因 |
| 留痕信息 | 合同版本、法源版本、模型与提示词版本、核验人及时间 |

## 执行步骤
1. 核对合同版本与类型，确认审查基准（范本 + 让步边界 + 我方立场）。
2. 按条款清单逐项抽取：主体、标的、价款与支付、履行期限、验收、违约、责任限制、保密、知识产权、争议解决、管辖、不可抗力、变更与解除。
3. 与范本逐项比对，标注差异性质与建议处置。
4. 对不利差异检索法条与制度依据，**逐条回源核验**；核验不通过的引用直接删除。
5. 生成风险标注清单与修改建议草稿。
6. 自校验：条款状态三类结论齐全；引用核验通过率 100%；差异项无遗漏静默。
7. 输出草稿并转具名律师核验。
8. 归档留痕，核验结论回流为评估样本。

## 质量标准（DoD）
- 一票否决：存在未回源核验的法条引用；使用了模型记忆提供条文；自主写入 CLM 正式字段或用印；审查了错误版本却未发现。
- 条款抽取完整性：条款清单覆盖率 100%，每项均有明确状态（不得为空白）。
- 引用核验通过率 100%，核验状态不得出现"未核验"。
- 差异项 100% 标注性质与建议处置。
- 未能审查的部分必须显式声明，不得静默跳过。

## 常见失败与处理
| 失败模式 | 表现 | 处置 |
|---|---|---|
| 法条幻觉 | 生成不存在的法条编号或条文内容 | 强制回源核验；核验不通过即删除该条及依赖结论 |
| 案号错配 | 案号真实但案情描述与裁判文书不符 | 逐条比对裁判要旨原文；不符即删除 |
| 版本错审 | 审查了过期版本 | 版本号前置校验；与 CLM 中最新版本比对 |
| 静默遗漏 | 条款清单中某项未抽取也未声明 | 强制输出三态结论；空白项阻断交付 |
| 立场错判 | 未考虑我方立场，给出中性但无用的意见 | 立场为必填输入；差异性质按立场重新标注 |
| 效力过时 | 引用已废止条款 | 强制携带生效日期；废止条款禁止加载 |

## 示例
输入：
- 合同文本：《XX 服务合同》v3.2
- 合同类型：服务合同；适用法域：中国内地；我方立场：甲方（采购方）
- 内部范本：服务合同范本 v2.0；让步边界清单：责任限制上限不超过合同总额

输出（节选）：
| 条款类型 | 位置 | 核心内容 | 状态 |
|---|---|---|---|
| 责任限制 | 第 9.2 条 | 乙方责任上限为合同总额的 50% | 已抽取 |
| 知识产权 | — | — | 未发现（建议补充） |
| 不可抗力 | 第 12 条 | 约定不可抗力范围与通知期限 | 已抽取 |

| 差异项 | 与范本差异 | 性质 | 建议处置 |
|---|---|---|---|
| 责任限制 | 范本为 100%，本合同为 50% | 不利 | 建议谈判恢复至不低于 100%，或在价款中作对价调整 |

| 风险点 | 风险等级 | 依据（已核验） | 修改建议 |
|---|---|---|---|
| 责任上限过低 | 中 | [法条以法源库检索结果为准，核验状态：已核验] | 建议调整上限并明确除外情形 |

> 说明：上表中条款位置与比例为示例格式，实际必须以合同原文与法源库返回结果填充，不得沿用示例值。

4.3 Rollout Checklist

5. Summary

The AI Harness for the Legal direction is essentially a system that engineers "citation authenticity" into a mandatory, enforceable mechanism.

Three conclusions:

First, the one non-negotiable indicator for the Legal direction is the citation-verification pass rate, and it must be 100%. Other directions may tolerate 99%, but the Legal direction cannot. The reason is simple: missing a non-core clause in contract review causes limited loss, but placing a single fabricated case number in a representation opinion costs professional liability and harms the order of the judicial system. The California Court of Appeal has already explicitly required that "all legal citations be verified line by line by attorneys against the originals," and Chinese courts have, through database-included cases, established the rule that "submitting false cases that are not screened or verified carries responsibility" — these two rules from different jurisdictions point to the same engineering requirement.

Second, "AI involvement" should be proactively disclosed, not concealed. In the U.S. Fifth Circuit case, "deliberately concealing AI use" was an aggravating circumstance for discipline. Although Chinese courts have not separately aggravated concealment, the Tongzhou court explicitly wrote into its judgment that "artificial-intelligence models must not be allowed to generate or fabricate false information that disrupts judicial order." Harness should firmly attach an AI-participation disclosure and human-verification information to its output.

Third, retaining generation-process records is both a compliance requirement and a form of self-protection. In the AI face-swapping case, the defendant bore the adverse consequence of failure of proof because it "could not reproduce the AI generation process it claimed." This adjudicative logic can be fully transferred to the legal scenario: when an AI-assisted material sparks a dispute, whether the generation process can be reproduced (what materials, what prompts, what model version were used) will directly affect the determination of liability. Therefore retaining prompts, input materials, model versions, and parameters is not optional.

A limitation that must be faced squarely: many of the foreign cases cited in this direction come from media retellings, and the specific penalties and procedures are subject to the formal documents of the courts in each jurisdiction; the progress status of some Chinese cases (e.g., whether an appeal was filed or retrial entered) could not be verified one by one, and citations should note "subject to the final determination by the competent authority." In addition, the copyrightability of AI-generated output has not yet been fully unified in judicial practice, and the adjudicative development after these cases is still evolving.

Information-Gap Disclosure

The following data could not be corroborated by a reliable secondary source during the writing of this direction and has been marked for processing per the specification, to be supplemented:

1. The full original text of the judgment in the Tencent Dreamwriter case: the content cited here comes from media retellings by Legal Daily's website (fazhiwang), 36 Kr, and others; the full text of the judgment was not obtained, and it is marked [To be verified].
2. The specific case number and judgment date of the Beijing Internet Court AI face-swapping case: the report did not disclose the full case number; it is only known that the case was concluded by the Beijing Internet Court and the judgment has taken effect, and it is marked [To be filled].
3. The trial court and case number of the Wu v. a technology company AI face-swap APP case: The Paper's report did not fully disclose them; it is only known that it is a related Fuzhou Intermediate People's Court case, and it is marked [To be verified].
4. The adjudicative difference between the Tencent Dreamwriter case and the Beijing Internet Court computer-software-generated-output case: some views hold that the two cases differ in their findings, but a complete comparative analysis of the two cases was not obtained, and this direction's conclusion is marked "not yet fully unified in practice."
5. The official judgment-document numbers of the various foreign cases (Mata case, California cases, Mississippi case, etc.): all are media retellings, and the original document numbers were not obtained; they are marked [To be verified].
6. The case number and progress of the Xiaogan Dawu County Court, Hubei case: it is only found in reports by CCTV's "Rule of Law Online" (Fazhi Zaixian), which did not disclose the case number; it is marked [To be filled].

6. References

1. AI 判例造假有恃无恐，全球携手严打 — 法治日报，2026。http://www.lawnewscn.cn/wap/content/2026-08/26/content_9446481.html
2. 法治在线 | 律师用 AI 生成虚假案例被法院发现 将承担何种责任 — 央广网，2026。https://news.cnr.cn/native/gd/20260115/t20260115_527494281.shtml
3. 警惕 AI 生成虚假案例干扰司法审理 — 人民网，2026。https://yn.people.com.cn/BIG5/n2/2026/0121/c361322-41478254.html
4. 媒体关注 | 律师提交 AI 生成虚假案例被当庭训诫 — 宜昌长安网，2026。http://zfw.yichang.gov.cn/content-64133-965206-1.html
5. AI 作品著作权归谁？法院审判多起"首案"厘清侵权边界 — 法治网，2026。https://www.legaldaily.com.cn/index/content/2026-06/03/content_9400177.html
6. 你用 AI 写的文章，著作权到底归谁？— 天达共和律师事务所，2026。http://www.east-concord.com/zygd/Article/20266/ArticleContent_4652.html
7. 技术不是侵权"挡箭牌" 法院这样认定 AI"盗脸" — 经济参考网，2026。https://www.jjckb.cn/20260418/957b6bc4d3a34c27b9954fe4d240ae97/c.html
8. 《家事法庭》联动普法：你的声音肖像，AI 说了不算！— 澎湃新闻。https://m.thepaper.cn/newsDetail_forward_32953019
9. 人工智能写作领域第一案在深圳法院落槌首次确认 — 法制日报 / 36 氪。https://36kr.com/coop/uc/5283237.html
10. 关于印发《人工智能生成合成内容标识办法》的通知（国信办通字〔2025〕2 号）— 国家互联网信息办公室等四部门，2025。https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
11. 中央网信办部署开展"清朗·整治 AI 技术滥用"专项行动 — 中央网络安全和信息化委员会办公室，2025。https://www.cac.gov.cn/2025-04/30/c_1747719097461951.htm
12. Secure Software Development Practices for Generative AI and Dual-Use Foundation Models（NIST SP 800-218A）— NIST CSRC，2024。https://csrc.nist.gov/pubs/sp/800/218/a/final