附录 · 术语表
1. 使用说明
- 本表汇总本白皮书及全部调研文档中使用的术语与缩写，共 10 组 90 余条，是全库术语的唯一定义处。各文档正文首次出现缩写时仍需给出全称。
- 释义力求与调研库原文一致；凡释义中含有数字者，其可信度以原始文档的标注为准（ 者仍须视为待核实）。
- “主要出现位置”标注术语最集中出现的目录或文件，格式为“目录 / 文件”。
- 英文原文与缩写保留原文写法；同一概念在不同来源存在多种译名时，本表以“中文术语”栏为全库统一译名。
- 本白皮书未采用的说法与类比，见第 3 章，引用调研库时应避免混用。
2. 术语与缩写总表
图 2-1｜术语表核心结构：六层能力模型与概念定位
数据来源：基于本文分析绘制的示意图。
2.1. 核心概念与工程体系
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 人工智能驾驭层 | AI Harness | — | 位于基础模型与真实业务系统之间的工程化承载层，把模型的不确定性转化为工程可预期性；不提升模型智能 | 01-概述 / 01-介绍；04-白皮书 / 02-定义 |
| 测试夹具 | Test Harness | — | 软件工程中外部于被测软件的执行环境，让功能可被自动、可重复、可判定地验证；AI Harness 的词源 | 01-概述 / 01-介绍 |
| Harness 工程 | Harness Engineering | — | 设计与持续改进 Harness 的工程方法论；2026 年 2 月由 Hashimoto、OpenAI、martinfowler.com 先后命名 | 01-概述 / 02-发展历史 |
| 六层能力模型 | Six-layer Capability Model | L1~L6 | L1 上下文工程、L2 工具与执行、L3 编排与控制、L4 记忆与状态、L5 评估与观测、L6 治理与安全的统一分层框架 | 01-概述 / 01-介绍；04-白皮书 / 02-定义 |
| 三代架构演进 | Three Generations of Architecture | — | 提示词工程时代（约 2020—2023）→ 工具与编排时代（约 2023—2025）→ 运行时与评估时代（2025—至今）的统一分期 | 01-概述 / 02-发展历史 |
| 智能体 | Agent | — | 具备自主感知、记忆、决策、交互与执行能力的新型智能系统；运行在 Harness 之上的执行单元 | 全库 |
| 智能体框架 | Agent Framework | — | 提供构建块的中间件（如 LangGraph、AutoGen），主要覆盖 L2/L3，是 Harness 的子集 | 01-概述 / 01-介绍 |
| 智能体平台 | Agent Platform | — | Harness 的产品化封装，叠加 UI、多租户、计费等能力 | 01-概述 / 01-介绍；03-市场研究 / 02-AI-Agents组 |
| AI 集成开发环境 | AI IDE | — | Harness 在编码场景的垂直集成形态，如 Cursor、Claude Code | 03-市场研究 / 01-AI-IDE组 |
| 上下文窗口 | Context Window | — | 模型单次推理可见的 token 范围，类比计算机的 RAM | 01-概述 / 01-介绍 |
| 标准作业程序 | Standard Operating Procedure | SOP | 固化的任务执行流程，AGENTS.md 骨架中的必备章节 | 各组 AGENTS.md |
| 完成定义 | Definition of Done | DoD | 任务完成的可判定标准，SKILL.md 骨架中的必备章节 | 各组 SKILL.md |
| 基础模型 | Foundation Model | — | 大规模预训练模型，Harness 的被承载对象 | 全库 |
| 大语言模型 | Large Language Model | LLM | 以文本生成为核心能力的基础模型类别 | 全库 |
| 人工智能生成内容 | AI Generated Content | AIGC | 由 AI 参与生成的文本、图像、音频、视频等内容 | 02-行业赋能 / 07-创意产业组 |
| 上下文投毒 | Context Poisoning | — | 错误信息进入检索语料或长期记忆后持续传播的失效模式 | 02-行业赋能 / 05-知识协同组 |
| N×M 集成问题 | N×M Integration Problem | — | M 个模型应用对接 N 个数据源时无标准产生的乘积级适配成本；MCP 使其降为 N+M | 01-概述 / 02-发展历史 |
| 人为在环外 | Human-on-the-Loop | — | 人工不修改产物本身，而是修改产生产物的 Harness（与 in the loop 相对） | 01-概述 / 01-介绍 |
| 智能体互联 | Agent Interconnection | — | 智能体之间身份、描述、发现与交互的互联互通问题，中国国标体系的主题 | 01-概述 / 04-未来发展 |
2.2. 上下文工程与工具协议
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 上下文工程 | Context Engineering | — | 在 LLM 推理期间策展并维护最优 token 集合的一整套策略；六层模型的 L1 | 01-概述 / 01-介绍 |
| 上下文腐烂 | Context Rot | — | 上下文变长时性能渐进退化的现象，是性能梯度而非硬性悬崖 | 01-概述 / 01-介绍 |
| 注意力预算 | Attention Budget | — | 因注意力稀释效应，上下文是稀缺资源的概念框架 | 01-概述 / 01-介绍 |
| 上下文压缩 | Context Compaction | — | 对上下文就地摘要以保留连续性的机制，与 Context Reset 相对 | 01-概述 / 03-架构演进 |
| 上下文重置 | Context Reset | — | 清空上下文后重新注入交接产物，给干净白板 | 01-概述 / 03-架构演进 |
| 检索增强生成 | Retrieval-Augmented Generation | RAG | 以外部检索结果增强模型输入的技术；决定模型“看到什么”的核心手段 | 02-行业赋能 / 05-知识协同组 |
| 情境化检索 | Contextual Retrieval | — | Anthropic 提出的检索优化方法，检索失败率自 5.7% 降至叠加重排后的 1.9%（A 级） | 02-行业赋能 / 05-知识协同组 |
| 图检索增强生成 | GraphRAG | — | 以知识图谱为检索结构的 RAG 变体 | 02-行业赋能 / 05-知识协同组 |
| 知识图谱 | Knowledge Graph | KG | 以实体与关系组织知识的结构化网络，可提升复杂查询准确率 | 02-行业赋能 / 05-知识协同组 |
| 权限感知检索 | Permission-aware Retrieval | — | 在检索层而非生成后做权限过滤的工程要求 | 02-行业赋能 / 05-知识协同组 |
| 模型上下文协议 | Model Context Protocol | MCP | 把外部工具与数据源以标准接口暴露给智能体的开放协议；2024-11-25 发布，后捐入 AAIF | 01-概述 / 02-发展历史、03-架构演进 |
| 函数调用 | Function Calling | — | 通过 JSON Schema 描述函数供模型调用的 API 契约；2023-06-13 产品化 | 01-概述 / 02-发展历史 |
| 智能体间协议 | Agent2Agent Protocol | A2A | Google 发布、后捐入 Linux Foundation 的智能体间协作协议，v1.0 于 2026 年 3 月发布 | 01-概述 / 02-发展历史 |
| 智能体指令文件 | AGENTS.md | — | 指导编码智能体的开源格式文件，距离被编辑文件最近者生效；被 60,000+ 项目采用后捐入 AAIF | 全库 |
| 技能文件 | SKILL.md | — | Agent Skills 的最小单元文件，含 name 与 description 必填元数据 | 01-概述 / 03-架构演进；02-行业赋能 / 07-创意产业组 |
| 智能体技能 | Agent Skills | — | 2025-10-16 发布、2025-12-18 转为开放标准的技能封装机制 | 01-概述 / 02-发展历史 |
| 渐进式披露 | Progressive Disclosure | — | Skills 的三级加载机制：先元数据、再指令全文、再按需读取捆绑资源 | 02-行业赋能 / 07-创意产业组 |
| 计算机使用 | Computer Use | — | 让模型直接操作图形界面的能力，Anthropic 2024-10-22 公开 beta | 01-概述 / 02-发展历史 |
| KV 缓存复用 | KV Cache Reuse | — | 复用键值缓存以降低重复前缀计算成本的上下文工程手段 | 01-概述 / 01-介绍 |
| 提示词装配 | Prompt Assembly | — | 按优先级将指令、检索结果与示例组装为最终提示词的过程 | 01-概述 / 01-介绍 |
| 分层指令文件 | Layered Instruction Files | — | 按目录层级就近放置指令文件并以 @import 复用的上下文组织方式 | 02-行业赋能 / 03-软件工程组 |
2.3. 编排、记忆与评估观测
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 智能体循环 | Agent Loop | — | “推理—行动—观察”交替的主循环，源于 ReAct 范式 | 01-概述 / 01-介绍 |
| 计划-执行 | Plan-and-Execute | — | 先全局规划再分步执行的编排形态 | 01-概述 / 01-介绍 |
| 子智能体 | Sub-agent | — | 由主智能体派发的独立上下文执行单元，只回传浓缩摘要（官方参数 1,000~2,000 tokens） | 01-概述 / 01-介绍；02-行业赋能 / 06-数据科学组 |
| 交接 | Handoff | — | 多智能体之间任务与上下文的移交协议 | 02-行业赋能 / 03-软件工程组 |
| 编排者-执行者 | Orchestrator-Worker | — | 一个编排智能体调度多个执行智能体的架构，核心价值是上下文隔离 | 02-行业赋能 / 06-数据科学组 |
| 事件钩子 | Hooks | — | 在工具调用前后做确定性拦截的事件机制 | 02-行业赋能 / 03-软件工程组 |
| 有向无环图 | Directed Acyclic Graph | DAG | 确定性工作流的图结构编排形态 | 01-概述 / 01-介绍 |
| 会话存储 | Session Store | — | L4 层的会话态持久化组件 | 01-概述 / 01-介绍 |
| 情景记忆 / 语义记忆 | Episodic / Semantic Memory | — | L4 层两类长期记忆：经历类与知识类 | 01-概述 / 01-介绍 |
| 检查点 | Checkpoint | — | 任务状态的可恢复快照，需包含工具副作用才是完整的 | 01-概述 / 01-介绍 |
| 持久化执行 | Durable Execution | — | 中断后可从断点恢复的工作流执行引擎能力 | 02-行业赋能 / 05-知识协同组 |
| 人在回路 | Human-in-the-Loop | HITL | 人工参与任务流程的机制；断点位置设计决定其是否流于形式 | 全库 |
| 沙箱 | Sandbox | — | 限定动作作用范围与副作用边界的隔离执行环境；Anthropic 实证权限提示减少 84%（A 级） | 01-概述 / 01-介绍 |
| 轨迹 | Trace | — | 智能体执行全过程的事件与调用记录，可观测性与可复盘性的基础 | 01-概述 / 01-介绍 |
| 可观测性 | Observability | — | 对系统内部状态的度量与追踪能力，属 L5；与审计（L6）目的不同 | 01-概述 / 01-介绍 |
| 评估集 | Evaluation Set / Eval Set | — | 用于判定产出质量的标准化任务集合 | 01-概述 / 01-介绍 |
| 黄金数据集 | Golden Dataset / Golden Set | — | 组织内部固化的标准评测集，用于弥补公共基准缺位 | 02-行业赋能 / 05-知识协同组 |
| 回归集 | Regression Set | — | 每次提示词、模型或工具变更后必须全量重跑的固化任务集 | 02-行业赋能 / 03-软件工程组 |
| 大模型作为评审 | LLM-as-a-Judge | — | 用另一个 LLM 充当评分器；须固定模型版本与温度并记录 | 02-行业赋能 / 03-软件工程组 |
| 护栏 | Guardrails | — | 对模型输入输出施加的自动化约束机制，属 L6 | 01-概述 / 01-介绍 |
| 预算护栏 | Budget Guardrail | — | 对单次任务与周期成本的硬上限约束 | 01-概述 / 01-介绍 |
| 服务等级目标 | Service Level Objective | SLO | 可靠性目标值，DevOps/SRE 方向的度量锚点 | 02-行业赋能 / 03-软件工程组 |
| 服务等级协议 | Service Level Agreement | SLA | 对外的服务等级承诺 | 02-行业赋能 / 03-软件工程组 |
| 平均修复时间 | Mean Time To Repair | MTTR | 故障恢复时长指标，SRE 与安全方向的核心指标 | 02-行业赋能 / 03-软件工程组、08-风险合规组 |
| 平均检测时间 | Mean Time To Detect | MTTD | 威胁与故障的检测时长指标 | 02-行业赋能 / 08-风险合规组 |
2.4. 基准与评测
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 软件工程基准 | SWE-bench | — | 2023-10-10 发布的 GitHub 真实issue 修复基准：2,294 题、12 个 Python 仓库，ICLR 2024 Oral；确立程序化判分原则 | 01-概述 / 02-发展历史 |
| 人工校验子集 | SWE-bench Verified | — | 2024-08-13 发布的 500 题人工校验子集；2026 年前沿系统突破 80% | 01-概述 / 02-发展历史 |
| 高难度变体 | SWE-bench Pro | — | 多文件、长时程的更难变体；与 Verified 数字不可直接比较（任务分布不同） | 02-行业赋能 / 03-软件工程组 |
| 终端基准 | Terminal-Bench | — | Stanford + Laude Institute 2025 年发布的终端任务基准；方法论声明“排行榜排的是系统而非模型” | 01-概述 / 01-介绍 |
| 自由职业工程基准 | SWE-Lancer | — | 以真实自由职业工程任务与报酬计价的基准 | 02-行业赋能 / 03-软件工程组 |
| 开放世界基准 | OSWorld | — | 在真实操作系统环境中评测计算机操作能力的基准 | 03-市场研究 / 02-AI-Agents组 |
| 抽象推理基准 | Abstraction and Reasoning Corpus - AGI | ARC-AGI | 考察抽象推理能力的基准系列；引用成绩须注明推理力度配置 | 02-行业赋能 / 03-软件工程组 |
| 技能扩展性基准 | SkillsBench | — | 评测技能机制扩展性的基准 | 02-行业赋能 / 03-软件工程组 |
| 评测口径 | Evaluation Protocol / Configuration | — | 推理力度、工具可用性、步数上限等评测配置；引用数字必须带四要素（模型版本、榜单、日期、配置） | 02-行业赋能 / 03-软件工程组 |
| 污染检测 | Contamination Check | — | 检查评测任务是否进入模型训练语料的核查动作 | 02-行业赋能 / 03-软件工程组 |
2.5. 行业域术语：硬件研发
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 电子设计自动化 | Electronic Design Automation | EDA | 芯片设计、仿真、验证与物理实现的工具链总称；L2 层不可替代的执行器 | 02-行业赋能 / 04-硬件研发组 |
| 寄存器传输级 | Register Transfer Level | RTL | 以寄存器与组合逻辑描述电路行为的设计抽象层级 | 02-行业赋能 / 04-硬件研发组 |
| 通用验证方法学 | Universal Verification Methodology | UVM | 基于 SystemVerilog 的标准化验证平台方法学（IEEE 1800.2-2020） | 02-行业赋能 / 04-硬件研发组 |
| 形式验证 | Formal Verification | — | 用数学方法穷尽证明设计满足给定属性，不依赖测试向量 | 02-行业赋能 / 04-硬件研发组 |
| 逻辑等价性检查 | Logic Equivalence Checking | LEC | 证明优化前后网表与黄金 RTL 功能等价的形式化检查 | 02-行业赋能 / 04-硬件研发组 |
| 可测性设计 | Design for Test | DFT | 插入扫描链、内建自测试等结构以提升可测性的方法 | 02-行业赋能 / 04-硬件研发组 |
| 自动测试设备 | Automated Test Equipment | ATE | 对芯片施加测试向量并判定合格与否的量产测试机台 | 02-行业赋能 / 04-硬件研发组 |
| 流片 | Tape-out | — | 将最终版图数据交付代工厂制造的关键节点，不可逆，代价极高 | 02-行业赋能 / 04-硬件研发组 |
| 芯粒 | Chiplet | — | 以裸片为粒度组合封装的集成方式；先进封装的核心形态 | 02-行业赋能 / 04-硬件研发组 |
| 通用芯粒互连标准 | Universal Chiplet Interconnect Express | UCIe | Die-to-Die 互连开放标准，3.0 版于 2025-08-05 发布 | 02-行业赋能 / 04-硬件研发组 |
| 覆盖率收敛 | Coverage Closure | — | 使功能覆盖率与代码覆盖率达到签核目标的过程 | 02-行业赋能 / 04-硬件研发组 |
| 模型算力利用率 | Model FLOPs Utilization | MFU | 实际有效算力与理论峰值之比，训练集群核心效率指标 | 02-行业赋能 / 04-硬件研发组 |
| 有效训练时间比率 | Effective Training Time Ratio | ETTR | 有效训练时间占总占用时间的比例，反映容错与恢复能力 | 02-行业赋能 / 04-硬件研发组 |
| 电源使用效率 | Power Usage Effectiveness | PUE | 数据中心总能耗与 IT 设备能耗之比，越接近 1 越优 | 02-行业赋能 / 04-硬件研发组 |
2.6. 行业域术语：数据与科学计算
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 高性能计算 | High Performance Computing | HPC | 以超算与并行作业为载体的计算形态；调度器是编排层，智能体不直接接触调度器 | 02-行业赋能 / 06-数据科学组 |
| 科学智能 / AI for Science | AI for Science | AI4S | 以 AI 支撑科学发现的方向；ground truth 是物理定律与实验验证而非人工标注 | 02-行业赋能 / 06-数据科学组 |
| 可复现性四层 | Repeatability / Replicability / Reproducibility / Reusability | — | ACM 对可复现性的四层术语定义，用作 L4 记忆结构设计的验收口径 | 02-行业赋能 / 06-数据科学组 |
| FAIR 原则 | Findable, Accessible, Interoperable, Reusable | — | 科研数据管理国际公认基准（2016） | 02-行业赋能 / 06-数据科学组 |
| 静默错误 | Silent Error | — | 管道不崩溃但结果错误——数据科学域最本质的风险形态 | 02-行业赋能 / 06-数据科学组 |
| 血缘 | Data Lineage | — | 数据从源到结果的依赖链路；OpenLineage 为其开放规范 | 02-行业赋能 / 06-数据科学组 |
| 有效效率得分 | Valid Efficiency Score | VES | 除执行准确率外考察查询效率的评估指标 | 02-行业赋能 / 06-数据科学组 |
2.7. 行业域术语：流程与知识协同
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 业务流程建模与标注 | Business Process Model and Notation | BPMN | OMG 业务流程建模标准（2.0.2，等同 ISO/IEC 19510:2013） | 02-行业赋能 / 05-知识协同组 |
| 决策模型与标注 | Decision Model and Notation | DMN | OMG 决策逻辑建模标准，与 BPMN 配套 | 02-行业赋能 / 05-知识协同组 |
| 案例管理建模与标注 | Case Management Model and Notation | CMMN | 表达知识工作者临时决策的 OMG 标准 | 02-行业赋能 / 05-知识协同组 |
| 以对象为中心的事件日志 | Object-Centric Event Log | OCEL | 流程挖掘的对象中心事件日志标准（2.0） | 02-行业赋能 / 05-知识协同组 |
| 流程挖掘 | Process Mining | — | 从事件日志还原实际流程并与规范模型比对的技术 | 02-行业赋能 / 05-知识协同组 |
| 业务流程管理 | Business Process Management | BPM | 决定业务按什么规则流转、是否合规、偏差能否检出的管理域 | 02-行业赋能 / 05-知识协同组 |
| 知识管理体系 | Knowledge Management Systems | — | ISO 30401:2018 确立的知识管理要求类标准 | 02-行业赋能 / 05-知识协同组 |
| 数字信号处理 | Digital Signal Processing | DSP | 信号处理技术领域（通用缩写） | 01-概述 / 参数卡缩写表 |
| 工件 | Artifact | — | 任务产出的可版本化对象（文档、模型、代码变更等） | 01-概述 / 01-介绍 |
2.8. 行业域术语：创意与内容产业
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 虚拟试穿 | Virtual Try-On | VTON | 图像生成方向以目标人物试穿目标服饰的能力 | 03-市场研究 / 03-AI-图像组 |
| 跨帧注意力 | Cross-frame Attention | — | 视频生成中对连续帧整体建模以抑制角色漂移的机制 | 02-行业赋能 / 07-创意产业组 |
| 镜头状态机 | Shot State Machine | — | 以镜头为单位记录生成状态、版本与校验结果的状态机 | 02-行业赋能 / 07-创意产业组 |
| 品牌资产库 | Brand Asset Library | — | 承载 VI、色卡、字体、角色 IP 等跨项目复用资产的库 | 02-行业赋能 / 07-创意产业组 |
| 角色设定表 | Character Sheet | — | 记录角色三视图、色卡、表情集的锚定上下文 | 02-行业赋能 / 07-创意产业组 |
| 显式标识 | Explicit Label | — | 在用户可感知位置添加的生成合成内容提示标识 | 02-行业赋能 / 07-创意产业组 |
| 隐式标识 | Implicit Label | — | 嵌入文件元数据的生成合成内容属性信息 | 02-行业赋能 / 07-创意产业组 |
| 资产管理壳 | Asset Administration Shell | AAS | 工业 4.0 组件的数字化表达（IEC 63278-1:2023） | 02-行业赋能 / 07-创意产业组 |
| 数字主线 | Digital Thread | — | 贯穿产品全生命周期的数据链路（ISO 23247-5:2026） | 02-行业赋能 / 07-创意产业组 |
| 制造运行管理 | Manufacturing Operations Management | MOM | IEC 62264（ISA-95）五层模型中的 L3 交界层 | 02-行业赋能 / 07-创意产业组 |
2.9. 治理、安全与合规
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 基于角色的访问控制 | Role-Based Access Control | RBAC | 按角色分配权限的权限模型，L6 核心组件 | 01-概述 / 01-介绍 |
| 漏洞概念证明 | Proof of Vulnerability | PoV | 以可复现的崩溃确认漏洞的判定方式（AIxCC 范式核心） | 02-行业赋能 / 08-风险合规组 |
| AI 网络攻防挑战赛 | AI Cyber Challenge | AIxCC | DARPA 主办的自主漏洞挖掘与修补竞赛，确立“模型编排、工具判定”范式 | 02-行业赋能 / 08-风险合规组 |
| AI 物料清单 | AI Bill of Materials | AI BOM | 对组织内 AI 资产的台账化管理 | 02-行业赋能 / 08-风险合规组 |
| 大模型应用十大风险 | OWASP Top 10 for LLM Applications | — | OWASP 发布的 LLM 应用风险清单（2025 版：提示词注入、敏感信息泄露等） | 02-行业赋能 / 08-风险合规组 |
| 过度代理 | Excessive Agency | — | 给予智能体超出必要权限与自主性的风险项 | 02-行业赋能 / 03-软件工程组、08-风险合规组 |
| 职责分离 | Separation of Duties | SoD | 生成与审核不得共用同一身份与凭证的治理要求（GB/T 45654—2025 第 4.3.1 条） | 02-行业赋能 / 08-风险合规组 |
| 个人身份信息脱敏 | PII Scrubbing | — | 对个人身份信息的识别与脱敏处理 | 02-行业赋能 / 06-数据科学组、08-风险合规组 |
| 影子 AI | Shadow AI | — | 未经治理私自使用的 AI 工具，泄露事件的重要来源 | 02-行业赋能 / 08-风险合规组 |
| 模型风险管理 | Model Risk Management | — | SR 11-7 确立的开发、验证、治理三支柱框架 | 02-行业赋能 / 08-风险合规组 |
| 证据链 | Chain of Evidence / Chain of Custody | — | 使结论可还原到依据的留痕结构；风险合规组的交付物主线 | 02-行业赋能 / 08-风险合规组 |
| 审计轨迹 | Audit Trail | — | 谁在何时依据什么做了什么的不可篡改记录 | 02-行业赋能 / 08-风险合规组 |
2.10. 标准组织、监管文件与调研机构
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 智能体人工智能基金会 | Agentic AI Foundation | AAIF | Linux Foundation 旗下基金会（2025-12-09 成立），托管 MCP、goose、AGENTS.md | 01-概述 / 02-发展历史 |
| 智能体互联系列国标 | GB/Z 185—2026 | — | 中国《人工智能 智能体互联》系列国家标准化指导性技术文件，七部分闭环，2026 年上半年发布 | 01-概述 / 02-发展历史、04-未来发展 |
| 生成式人工智能服务管理暂行办法 | Interim Measures for Generative AI Services | — | 七部门令第 15 号，2023-08-15 施行；第十九条为证据链的法定表述 | 02-行业赋能 / 08-风险合规组 |
| 人工智能生成合成内容标识办法 | Measures for Labeling AI-Generated Content | — | 2025-09-01 施行，要求显式与隐式标识、日志留存不少于六个月 | 02-行业赋能 / 07-创意产业组 |
| 生成式人工智能服务安全基本要求 | GB/T 45654—2025 | — | 2025-11-01 实施的国家标准：违法不良信息超 5% 不采集、标注执行与审核人员分离等 | 02-行业赋能 / 08-风险合规组 |
| 人工智能管理体系 | ISO/IEC 42001:2023 | — | 全球首个可认证的 AI 管理体系标准，附录 A 含 38 项控制 | 02-行业赋能 / 03-软件工程组、08-风险合规组 |
| 人工智能风险管理框架 | AI Risk Management Framework | AI RMF | NIST 发布的自愿性风险管理框架（1.0，2023-01-26）；GenAI Profile 为其生成式 AI 补充 | 02-行业赋能 / 08-风险合规组 |
| 欧盟人工智能法案 | EU AI Act, Regulation (EU) 2024/1689 | — | 2024-08-01 生效的欧盟人工智能法规，高风险义务含日志留存、技术文件与人工监督 | 02-行业赋能 / 08-风险合规组 |
| 合规管理体系 | Compliance Management Systems | — | ISO 37301:2021，等同转化为 GB/T 35770—2022 | 02-行业赋能 / 08-风险合规组 |
| 人工智能评测研究机构 | Model Evaluation & Threat Research | METR | 发布资深开发者 AI 效率随机对照试验（实测慢 19%）的研究机构 | 02-行业赋能 / 03-软件工程组 |
| 开发者效能研究项目 | DevOps Research and Assessment | DORA | Google 旗下研究项目，发布年度软件交付效能报告 | 02-行业赋能 / 03-软件工程组 |
| 安全软件开发框架 | Secure Software Development Framework | SSDF | NIST SP 800-218A（v1.1，2024-10） | 02-行业赋能 / 03-软件工程组 |
| 智能制造能力成熟度 | Smart Manufacturing Capability Maturity | — | GB/T 39116-2020 确立的成熟度模型 | 02-行业赋能 / 07-创意产业组 |
2.11. 行业域术语：AI Infra 与具身智能
本节随 04-实践 章扩展为八大行业组而补入，覆盖 AI Infra 组与具身智能组的核心术语。
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 人工智能基础设施 | AI Infrastructure | AI Infra | 支撑模型训练与推理的算力、框架、调度与运维栈；在 Harness 六层模型中主要对应 L2 执行层的物理底座 | 02-行业赋能 / 01-AI-Infra组；04-白皮书 / 04-实践 |
| 推理服务 | Inference Serving | — | 将训练好的模型部署为可调用服务的工程环节，核心指标为时延、吞吐与单 token 成本 | 02-行业赋能 / 01-AI-Infra组；03-市场研究 / 06-AI-Infra组 |
| 训练框架 | Training Framework | — | 组织分布式训练过程的软件栈，负责并行策略、显存优化与容错恢复 | 02-行业赋能 / 01-AI-Infra组 |
| 推理框架 | Inference Framework | — | 面向部署侧优化模型执行的软件栈，负责算子融合、量化与批处理调度 | 02-行业赋能 / 01-AI-Infra组 |
| 智能运维 | AI Operations | AIOps | 以 AI 能力增强集群与服务的监控、告警、根因定位与自愈 | 02-行业赋能 / 01-AI-Infra组 |
| 具身智能 | Embodied AI | — | 具备物理载体的智能系统，通过感知—决策—控制闭环在真实环境中行动 | 02-行业赋能 / 02-具身智能组；03-市场研究 / 07-具身智能组 |
| 视觉—语言—动作模型 | Vision-Language-Action Model | VLA | 将视觉与语言输入直接映射为动作输出的具身智能模型范式 | 02-行业赋能 / 02-具身智能组；04-白皮书 / 08-发展展望 |
| 小脑（运动控制层） | Cerebellum / Motor Control Layer | — | 具身系统中负责高频实时运动控制与姿态稳定的层级，与负责高层规划的大脑层相对 | 02-行业赋能 / 02-具身智能组 |
| 执行器 | Actuator | — | 将控制信号转换为物理动作的部件，是具身系统 L2 执行层的物理末端 | 02-行业赋能 / 02-具身智能组 |
| 仿真到现实迁移 | Sim-to-Real Transfer | Sim2Real | 将在仿真环境中习得的策略迁移到真实物理系统的方法及其差距问题 | 02-行业赋能 / 02-具身智能组 |
| 算力规模 | Compute Scale | — | 以加速卡数量或算力当量衡量的训练资源体量；2026 年起成为具身智能公司的竞争壁垒之一 | 04-白皮书 / 08-发展展望 |
2.12. 版本维护与变更记录
本节随 2026-09-12 快照引入的快照级变更记录机制而补入。
| 中文术语 | 英文 | 缩写 | 释义 | 主要出现位置 |
|---|---|---|---|---|
| 版本快照 | Snapshot | — | 以日期命名的完整文档库副本（Harness-Reports-YYMMDD），当日增量在其上进行，历史快照保持只读 | 04-白皮书 / 11-增量维护白皮书 |
| 增量更新 | Incremental Update | — | 每日只增补新增与修订内容、不重写全库的更新方式 | 04-白皮书 / 11-增量维护白皮书 |
| 清单文件 | Manifest | — | 记录快照内每个文件摘要基线的机器可读文件（MANIFEST.md5），含全文件 MD5 与正文指纹 | 各快照根目录 |
| 消息摘要算法第 5 版 | Message-Digest Algorithm 5 | MD5 | 生成文件摘要的哈希算法；本工程取其前 8 位作为变更判据的可读标识 | 各快照根目录 |
| 正文指纹 | Text Fingerprint | — | 剔除 SVG 块、图注行与来源行并规范化连续空行后的内容 MD5，用于区分「正文变更」与「仅图表变更」 | 各快照根目录 |
| 变更记录 | Changelog | — | 相对上一快照的文件级变更说明，分列正文变更、仅图表变更、新增与删除 | 各快照根目录 |
| 信息截止 | Information Cutoff | — | 快照内容所覆盖的事实时间上界；晚于该日期的事项不得写入该快照 | 04-白皮书 / 11-增量维护白皮书 |
| 证据等级 | Evidence Grade | A/B/C | 来源可信度分级：A 为官方原文，B 为权威媒体报道，C 为一般来源；未获官方确认者另标 [待核实] | 全库 |
3. 本白皮书未采用的说法
以下说法在本工程语料中出现过，或在业界流传，但本白皮书不将其作为正式术语或论据。引用调研库时应避免混用。
- “AI 操作系统 / Agent OS”：仅作计算机四层栈类比使用，不是正式术语。类比可帮助理解，但不能承担论证功能。
- “MCP 是 AI 的 USB-C”：仅见于科普转述，白皮书正文不采用此类类比作论据。
- “Harness 层已有权威市场规模测算”：Harness 层自身市场规模无权威测算（
[待填写]）；现有生成式 AI 市场预测口径差异近一个数量级，白皮书不采用任何此类数字。 - “模型更强后 Harness 会消失”：白皮书立场为“Harness 会变轻、会改构，但不会消失”，依据为官方自陈与后训练耦合现象。
- “Agent Framework 就是 Harness”：不采用。Framework 中立、Harness 有主张，前者是后者的子集。
- “智能体即数字员工”等拟人化表述：不采用。本白皮书统一使用“智能体（Agent）”这一中国官方口径下的中性术语。
- “榜单高分即落地能力”：不采用。公开榜单只能校准能力天花板，不能预测内部收益（任务分布不同、口径不可比）。
- 未经核实的厂商内部数字：如 LangChain 仅改 Harness 的 52.8% 对 66.5%、Vercel 工具裁剪的 80% 对 100%、Claude Code 七周 350,000 DAU、Codex Auto-review 的 1/200 与 99% 等——白皮书引用时一律保留 标注，不作为确定事实。
- “Agent Harness”“AI Harness”“Agent Runtime”混用：本白皮书统一使用“AI Harness”指代该工程层；“Agent Harness”仅指编码等具体场景的实现形态；“Agent Runtime”仅指其运行时产品形态。
- “大模型即水电气”式的无限供给论断：不采用。成本、配额与算力约束（预算护栏）是 L6 的一等约束，非无限资源。
4. 总结
- 本表共收录 10 组术语，覆盖核心概念、上下文与工具协议、编排记忆评估、基准评测，以及硬件、数据、流程、创意、合规五个行业域和标准监管机构，与全库术语口径一致。
- 全部释义以调研库原文为准；含数字的释义继承原始文档的 标注。
- 第 3 章列出的“未采用说法”是本白皮书与部分网络流行叙事的边界，引用本白皮书时请以本表口径为准。
- 术语的完整论证与上下文，见 02-定义与各行业组文档；引用规范格式的全文，见 04-实践各组 AGENTS.md / SKILL.md。
5. 信息缺口声明
- “Agent Harness”术语首创出处：无一手首创文献，本表不设首创者条目。
- SkillsBench、Win-Hill 等较新基准的官方规范原文：仅在调研库转述中出现，本表释义仅作索引，具体评测协议以官方文档为准。
- ~~《人工智能生成合成内容标识方法》的 GB 编号~~：已确认（原缺口关闭）——GB 45438—2025，强制性国标，2025-02-28 发布、2025-09-01 实施，与《标识办法》配套（来源：国家标准全文公开系统 std.samr.gov.cn、TC260 官方文本）。本表已收录该编号。
- GB/T 39116-2020 的能力子域口径：存在两种口径，调研库统一采用 PTRM 四要素 / 8 能力域 / 20 能力子域并注明另一口径，本表不展开。
- DSP、TTM 等通用缩写在特定文档中的专用含义：以出现文档的上下文为准，本表仅收录其通用释义。
- 个别术语在不同组文档中的细微口径差异（如“交接 Handoff”在多智能体框架中的协议细节）：以各方向文档原文为准，本表取最大公约数释义。
6. 参考资料
- 01-概述 / 01-介绍 — 统一定义、六层能力模型与概念边界的源头文件。见
01-概述/01-介绍.md - 01-概述 / 02-发展历史 — 三代演进与关键里程碑。见
01-概述/02-发展历史.md - .workbuddy / PARAM_CARD.md — 跨文档一致性的唯一基准。见
.workbuddy/PARAM_CARD.md - 02-行业赋能 / 03-软件工程组 README — 软件工程方向术语与度量口径。见
02-行业赋能/03-软件工程组/README.md - 02-行业赋能 / 04-硬件研发组 README — 硬件域术语表。见
02-行业赋能/04-硬件研发组/README.md - 02-行业赋能 / 05-知识协同组 README — 流程与知识域术语及标准缺口。见
02-行业赋能/05-知识协同组/README.md - 02-行业赋能 / 06-数据科学组 README — 数据与科学域术语。见
02-行业赋能/06-数据科学组/README.md - 02-行业赋能 / 07-创意产业组 README — 内容产业域术语与标识合规。见
02-行业赋能/07-创意产业组/README.md - 02-行业赋能 / 08-风险合规组 README — 治理合规域术语与监管基座。见
02-行业赋能/08-风险合规组/README.md - SWE-bench 官方站 — Princeton NLP 等。https://www.swebench.com/
- Terminal-Bench 官方站 — Stanford / Laude Institute。https://www.tbench.ai/
- Model Context Protocol 官方站与规范 — MCP / AAIF。https://modelcontextprotocol.io/
- AGENTS.md 官方站点 — Linux Foundation / Agentic AI Foundation。https://agents.md/
- Introducing Agent Skills — Anthropic，2025-10-16。https://www.anthropic.com/news/skills
Appendix · Glossary
1. How to Read This Glossary
- This table consolidates the terms and abbreviations used in this whitepaper and across all research documents — 10 groups, 90+ entries — and is the single source of definition for terms across the entire library. Documents must still spell out a full name the first time an abbreviation appears in the body text.
- Definitions aim to stay faithful to the original text in the research library; wherever a definition contains numbers, its reliability follows the notation of the original document (those marked
[to be verified]must still be treated as unverified). - “Primary location” notes the directory or file where a term appears most concentrated, in the format “directory / file”.
- English originals and abbreviations retain their original form; when the same concept has multiple translations in different sources, this table uses the “Chinese term” column as the library-wide unified translation.
- Phrasings and analogies not adopted by this whitepaper are listed in Chapter 3; avoid mixing them in when citing the research library.
2. Master Table of Terms and Abbreviations
图 2-1｜术语表核心结构：六层能力模型与概念定位
数据来源：基于本文分析绘制的示意图。
2.1. Core Concepts and Engineering Systems
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| AI Harness | AI Harness | — | The engineered carrier layer between foundation models and real business systems that turns model uncertainty into engineering predictability; it does not make the model more intelligent | 01-Overview / 01-Introduction; 04-Whitepaper / 02-Definitions |
| Test Harness | Test Harness | — | In software engineering, the execution environment external to the software under test that lets functionality be verified automatically, repeatedly, and deterministically; the etymon of AI Harness | 01-Overview / 01-Introduction |
| Harness Engineering | Harness Engineering | — | The engineering methodology for designing and continuously improving a Harness; named successively by Hashimoto, OpenAI, and martinfowler.com in February 2026 | 01-Overview / 02-History |
| Six-layer Capability Model | Six-layer Capability Model | L1~L6 | A unified layered framework: L1 Context Engineering, L2 Tools & Execution, L3 Orchestration & Control, L4 Memory & State, L5 Evaluation & Observability, L6 Governance & Security | 01-Overview / 01-Introduction; 04-Whitepaper / 02-Definitions |
| Three Generations of Architecture | Three Generations of Architecture | — | A unified periodization: the prompt-engineering era (c. 2020–2023) → the tools & orchestration era (c. 2023–2025) → the runtime & evaluation era (2025–present) | 01-Overview / 02-History |
| Agent | Agent | — | A new kind of intelligent system capable of autonomous perception, memory, decision-making, interaction, and execution; the execution unit that runs on a Harness | Whole library |
| Agent Framework | Agent Framework | — | Middleware that provides building blocks (e.g. LangGraph, AutoGen), covering mainly L2/L3; a subset of a Harness | 01-Overview / 01-Introduction |
| Agent Platform | Agent Platform | — | The productized packaging of a Harness, adding UI, multi-tenancy, billing, and other capabilities | 01-Overview / 01-Introduction; 03-Market Research / 02-AI-Agents group |
| AI IDE | AI IDE | — | The vertically integrated form of a Harness in the coding scenario, e.g. Cursor, Claude Code | 03-Market Research / 01-AI-IDE group |
| Context Window | Context Window | — | The range of tokens visible to the model in a single inference pass, analogous to a computer's RAM | 01-Overview / 01-Introduction |
| Standard Operating Procedure | Standard Operating Procedure | SOP | A fixed task-execution procedure, a required section of the AGENTS.md skeleton | Per-group AGENTS.md |
| Definition of Done | Definition of Done | DoD | A determinable criterion for a task being complete, a required section of the SKILL.md skeleton | Per-group SKILL.md |
| Foundation Model | Foundation Model | — | A large pre-trained model; the entity carried by a Harness | Whole library |
| Large Language Model | Large Language Model | LLM | The class of foundation models whose core capability is text generation | Whole library |
| AI Generated Content | AI Generated Content | AIGC | Text, images, audio, video, and other content generated with AI participation | 02-Industry Enablement / 07-Creative Industry group |
| Context Poisoning | Context Poisoning | — | A failure mode in which wrong information keeps propagating once it enters retrieval corpora or long-term memory | 02-Industry Enablement / 05-Knowledge Collaboration group |
| N×M Integration Problem | N×M Integration Problem | — | The multiplicative adaptation cost, absent a standard, of connecting M model applications to N data sources; MCP reduces it to N+M | 01-Overview / 02-History |
| Human-on-the-Loop | Human-on-the-Loop | — | Humans do not edit the artifact itself but the Harness that produces it (as opposed to in the loop) | 01-Overview / 01-Introduction |
| Agent Interconnection | Agent Interconnection | — | The interoperability problem of identity, description, discovery, and interaction among agents; the theme of China's national-standard system | 01-Overview / 04-Future Outlook |
2.2. Context Engineering and Tool Protocols
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Context Engineering | Context Engineering | — | The whole set of strategies for curating and maintaining the optimal token set during LLM inference; L1 of the six-layer model | 01-Overview / 01-Introduction |
| Context Rot | Context Rot | — | The phenomenon of gradually degrading performance as context grows; a performance gradient rather than a hard cliff | 01-Overview / 01-Introduction |
| Attention Budget | Attention Budget | — | The conceptual framework that treats context as a scarce resource due to attention-dilution effects | 01-Overview / 01-Introduction |
| Context Compaction | Context Compaction | — | In-place summarization of context to preserve continuity; in contrast to Context Reset | 01-Overview / 03-Architecture Evolution |
| Context Reset | Context Reset | — | Clearing the context and re-injecting the handoff artifacts to give a clean slate | 01-Overview / 03-Architecture Evolution |
| Retrieval-Augmented Generation | Retrieval-Augmented Generation | RAG | The technique of augmenting model input with external retrieval results; the core means of deciding what the model “sees” | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Contextual Retrieval | Contextual Retrieval | — | Anthropic’s retrieval-optimization method; lowers the retrieval failure rate from 5.7% to 1.9% after re-ranking (Grade A) | 02-Industry Enablement / 05-Knowledge Collaboration group |
| GraphRAG | GraphRAG | — | A RAG variant that uses a knowledge graph as its retrieval structure | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Knowledge Graph | Knowledge Graph | KG | A structured network that organizes knowledge by entities and relations; can improve accuracy on complex queries | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Permission-aware Retrieval | Permission-aware Retrieval | — | The engineering requirement to filter permissions at the retrieval layer rather than after generation | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Model Context Protocol | Model Context Protocol | MCP | An open protocol exposing external tools and data sources to agents through a standard interface; released 2024-11-25, later donated to AAIF | 01-Overview / 02-History, 03-Architecture Evolution |
| Function Calling | Function Calling | — | An API contract describing functions in JSON Schema for the model to call; productized 2023-06-13 | 01-Overview / 02-History |
| Agent2Agent Protocol | Agent2Agent Protocol | A2A | An inter-agent collaboration protocol released by Google and later donated to the Linux Foundation; v1.0 released March 2026 | 01-Overview / 02-History |
| AGENTS.md | AGENTS.md | — | An open-source format file that guides coding agents; the one closest to the file being edited takes effect; donated to AAIF after adoption by 60,000+ projects | Whole library |
| SKILL.md | SKILL.md | — | The minimal unit file of Agent Skills, with required name and description metadata | 01-Overview / 03-Architecture Evolution; 02-Industry Enablement / 07-Creative Industry group |
| Agent Skills | Agent Skills | — | A skill-packaging mechanism released 2025-10-16 and turned into an open standard 2025-12-18 | 01-Overview / 02-History |
| Progressive Disclosure | Progressive Disclosure | — | Skills’ three-level loading mechanism: metadata first, then full instructions, then bundled resources read on demand | 02-Industry Enablement / 07-Creative Industry group |
| Computer Use | Computer Use | — | The ability to let a model operate graphical interfaces directly; public beta by Anthropic 2024-10-22 | 01-Overview / 02-History |
| KV Cache Reuse | KV Cache Reuse | — | A context-engineering technique that reuses key-value caches to cut the cost of recomputing repeated prefixes | 01-Overview / 01-Introduction |
| Prompt Assembly | Prompt Assembly | — | The process of assembling instructions, retrieval results, and examples into the final prompt by priority | 01-Overview / 01-Introduction |
| Layered Instruction Files | Layered Instruction Files | — | A context-organization approach that places instruction files close to directory levels and reuses them via @import | 02-Industry Enablement / 03-Software Engineering group |
2.3. Orchestration, Memory, and Evaluation & Observability
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Agent Loop | Agent Loop | — | The main “reason–act–observe” alternating loop, originating from the ReAct paradigm | 01-Overview / 01-Introduction |
| Plan-and-Execute | Plan-and-Execute | — | An orchestration form that plans globally first, then executes step by step | 01-Overview / 01-Introduction |
| Sub-agent | Sub-agent | — | An independent-context execution unit dispatched by a main agent that returns only a condensed summary (official parameter 1,000–2,000 tokens) | 01-Overview / 01-Introduction; 02-Industry Enablement / 06-Data Science group |
| Handoff | Handoff | — | The protocol for transferring tasks and context between multiple agents | 02-Industry Enablement / 03-Software Engineering group |
| Orchestrator-Worker | Orchestrator-Worker | — | An architecture in which one orchestrator agent schedules multiple worker agents; its core value is context isolation | 02-Industry Enablement / 06-Data Science group |
| Hooks | Hooks | — | An event mechanism for deterministic interception before and after tool calls | 02-Industry Enablement / 03-Software Engineering group |
| Directed Acyclic Graph | Directed Acyclic Graph | DAG | A graph-structured orchestration form for deterministic workflows | 01-Overview / 01-Introduction |
| Session Store | Session Store | — | The L4-layer component that persists session state | 01-Overview / 01-Introduction |
| Episodic / Semantic Memory | Episodic / Semantic Memory | — | The two kinds of long-term memory at the L4 layer: experience-based and knowledge-based | 01-Overview / 01-Introduction |
| Checkpoint | Checkpoint | — | A recoverable snapshot of task state; complete only when it includes tool side effects | 01-Overview / 01-Introduction |
| Durable Execution | Durable Execution | — | The workflow-execution engine capability of resuming from a checkpoint after interruption | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Human-in-the-Loop | Human-in-the-Loop | HITL | The mechanism of human participation in the task flow; where breakpoints are placed determines whether it is merely ceremonial | Whole library |
| Sandbox | Sandbox | — | An isolated execution environment that bounds the scope of actions and the boundaries of side effects; Anthropic empirically measured an 84% reduction in permission prompts (Grade A) | 01-Overview / 01-Introduction |
| Trace | Trace | — | The record of events and calls across an agent’s whole execution; the basis for observability and retrospective review | 01-Overview / 01-Introduction |
| Observability | Observability | — | The ability to measure and trace internal system state; part of L5; distinct in purpose from auditing (L6) | 01-Overview / 01-Introduction |
| Evaluation Set / Eval Set | Evaluation Set / Eval Set | — | A standardized set of tasks used to judge output quality | 01-Overview / 01-Introduction |
| Golden Dataset / Golden Set | Golden Dataset / Golden Set | — | A standard evaluation set fixed within an organization, used to fill gaps left by public benchmarks | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Regression Set | Regression Set | — | The fixed task set that must be rerun in full after every prompt, model, or tool change | 02-Industry Enablement / 03-Software Engineering group |
| LLM-as-a-Judge | LLM-as-a-Judge | — | Using another LLM as the scorer; the model version and temperature must be fixed and recorded | 02-Industry Enablement / 03-Software Engineering group |
| Guardrails | Guardrails | — | Automated constraint mechanisms applied to model input and output; part of L6 | 01-Overview / 01-Introduction |
| Budget Guardrail | Budget Guardrail | — | A hard upper-bound constraint on per-task and period cost | 01-Overview / 01-Introduction |
| Service Level Objective | Service Level Objective | SLO | Reliability target value; the measurement anchor in DevOps/SRE | 02-Industry Enablement / 03-Software Engineering group |
| Service Level Agreement | Service Level Agreement | SLA | An external commitment to service-level targets | 02-Industry Enablement / 03-Software Engineering group |
| Mean Time To Repair | Mean Time To Repair | MTTR | Fault-recovery duration metric; a core metric in SRE and security | 02-Industry Enablement / 03-Software Engineering, 08-Risk & Compliance group |
| Mean Time To Detect | Mean Time To Detect | MTTD | The metric for how long threats and faults take to detect | 02-Industry Enablement / 08-Risk & Compliance group |
2.4. Benchmarks and Evaluation
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Software Engineering Benchmark | SWE-bench | — | A GitHub real-issue-fixing benchmark released 2023-10-10: 2,294 problems across 12 Python repos, ICLR 2024 Oral; established the principle of programmatic grading | 01-Overview / 02-History |
| Human-Verified Subset | SWE-bench Verified | — | A 500-problem human-verified subset released 2024-08-13; frontier systems broke past 80% by 2026 | 01-Overview / 02-History |
| Hard Variant | SWE-bench Pro | — | A harder multi-file, long-horizon variant; its numbers cannot be directly compared with Verified (different task distributions) | 02-Industry Enablement / 03-Software Engineering group |
| Terminal Benchmark | Terminal-Bench | — | A terminal-task benchmark released in 2025 by Stanford + Laude Institute; its methodology declares that “leaderboards rank systems, not models” | 01-Overview / 01-Introduction |
| Freelance Engineering Benchmark | SWE-Lancer | — | A benchmark priced by real freelance engineering tasks and payouts | 02-Industry Enablement / 03-Software Engineering group |
| Open-World Benchmark | OSWorld | — | A benchmark that evaluates computer-operation ability in a real operating-system environment | 03-Market Research / 02-AI-Agents group |
| Abstract-Reasoning Benchmark | Abstraction and Reasoning Corpus - AGI | ARC-AGI | A benchmark series for abstract-reasoning ability; citations of scores must state the reasoning-effort configuration | 02-Industry Enablement / 03-Software Engineering group |
| Skill-Scalability Benchmark | SkillsBench | — | A benchmark that evaluates the scalability of the skill mechanism | 02-Industry Enablement / 03-Software Engineering group |
| Evaluation Protocol / Configuration | Evaluation Protocol / Configuration | — | Evaluation configuration such as reasoning effort, tool availability, and step limits; any cited number must carry the four elements (model version, leaderboard, date, configuration) | 02-Industry Enablement / 03-Software Engineering group |
| Contamination Check | Contamination Check | — | The verification action of checking whether evaluation tasks have entered the model training corpus | 02-Industry Enablement / 03-Software Engineering group |
2.5. Industry-Domain Terms: Hardware R&D
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Electronic Design Automation | Electronic Design Automation | EDA | The umbrella term for the toolchain of chip design, simulation, verification, and physical implementation; an irreplaceable executor at the L2 layer | 02-Industry Enablement / 04-Hardware R&D group |
| Register Transfer Level | Register Transfer Level | RTL | The design-abstraction level describing circuit behavior with registers and combinational logic | 02-Industry Enablement / 04-Hardware R&D group |
| Universal Verification Methodology | Universal Verification Methodology | UVM | The standardized SystemVerilog-based verification-platform methodology (IEEE 1800.2-2020) | 02-Industry Enablement / 04-Hardware R&D group |
| Formal Verification | Formal Verification | — | Exhaustively proving by mathematical methods that a design satisfies given properties, without relying on test vectors | 02-Industry Enablement / 04-Hardware R&D group |
| Logic Equivalence Checking | Logic Equivalence Checking | LEC | A formal check proving that the netlist before and after optimization is functionally equivalent to the golden RTL | 02-Industry Enablement / 04-Hardware R&D group |
| Design for Test | Design for Test | DFT | Methods that insert scan chains, built-in self-test, and other structures to improve testability | 02-Industry Enablement / 04-Hardware R&D group |
| Automated Test Equipment | Automated Test Equipment | ATE | Production test machines that apply test vectors to chips and judge pass/fail | 02-Industry Enablement / 04-Hardware R&D group |
| Tape-out | Tape-out | — | The critical, irreversible, and very costly node at which final layout data is handed to the foundry for manufacturing | 02-Industry Enablement / 04-Hardware R&D group |
| Chiplet | Chiplet | — | An integration method that packages at the bare-die granularity; the core form of advanced packaging | 02-Industry Enablement / 04-Hardware R&D group |
| Universal Chiplet Interconnect Express | Universal Chiplet Interconnect Express | UCIe | An open Die-to-Die interconnect standard; version 3.0 released 2025-08-05 | 02-Industry Enablement / 04-Hardware R&D group |
| Coverage Closure | Coverage Closure | — | The process of bringing functional and code coverage up to signoff targets | 02-Industry Enablement / 04-Hardware R&D group |
| Model FLOPs Utilization | Model FLOPs Utilization | MFU | The ratio of actual effective compute to theoretical peak; a core efficiency metric for training clusters | 02-Industry Enablement / 04-Hardware R&D group |
| Effective Training Time Ratio | Effective Training Time Ratio | ETTR | The proportion of effective training time to total occupied time; reflects fault tolerance and recovery capability | 02-Industry Enablement / 04-Hardware R&D group |
| Power Usage Effectiveness | Power Usage Effectiveness | PUE | The ratio of a data center’s total energy use to its IT equipment energy use; the closer to 1, the better | 02-Industry Enablement / 04-Hardware R&D group |
2.6. Industry-Domain Terms: Data and Scientific Computing
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| High Performance Computing | High Performance Computing | HPC | A computing form carried by supercomputers and parallel jobs; the scheduler is the orchestration layer, and agents do not touch the scheduler directly | 02-Industry Enablement / 06-Data Science group |
| AI for Science | AI for Science | AI4S | The direction of using AI to support scientific discovery; ground truth is physical laws and experimental verification, not human annotation | 02-Industry Enablement / 06-Data Science group |
| Four R’s of Reproducibility | Repeatability / Replicability / Reproducibility / Reusability | — | ACM’s four-tier terminology for reproducibility, used as the acceptance criterion for L4 memory-structure design | 02-Industry Enablement / 06-Data Science group |
| FAIR Principles | Findable, Accessible, Interoperable, Reusable | — | The internationally recognized baseline for research data management (2016) | 02-Industry Enablement / 06-Data Science group |
| Silent Error | Silent Error | — | A pipeline that does not crash but produces wrong results — the most fundamental risk shape in the data-science domain | 02-Industry Enablement / 06-Data Science group |
| Data Lineage | Data Lineage | — | The dependency chain of data from source to result; OpenLineage is its open specification | 02-Industry Enablement / 06-Data Science group |
| Valid Efficiency Score | Valid Efficiency Score | VES | An evaluation metric measuring query efficiency in addition to execution accuracy | 02-Industry Enablement / 06-Data Science group |
2.7. Industry-Domain Terms: Process and Knowledge Collaboration
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Business Process Model and Notation | Business Process Model and Notation | BPMN | OMG’s business-process modeling standard (2.0.2, equivalent to ISO/IEC 19510:2013) | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Decision Model and Notation | Decision Model and Notation | DMN | OMG’s standard for modeling decision logic, paired with BPMN | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Case Management Model and Notation | Case Management Model and Notation | CMMN | An OMG standard expressing knowledge workers’ ad-hoc decisions | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Object-Centric Event Log | Object-Centric Event Log | OCEL | The object-centric event-log standard for process mining (2.0) | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Process Mining | Process Mining | — | The technique of reconstructing actual processes from event logs and comparing them with normative models | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Business Process Management | Business Process Management | BPM | The management domain deciding by what rules a business flows, whether it is compliant, and whether deviations can be detected | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Knowledge Management Systems | Knowledge Management Systems | — | The requirements-type standard for knowledge management established by ISO 30401:2018 | 02-Industry Enablement / 05-Knowledge Collaboration group |
| Digital Signal Processing | Digital Signal Processing | DSP | The signal-processing technical field (general abbreviation) | 01-Overview / PARAM_CARD abbreviation table |
| Artifact | Artifact | — | A versionable object produced by a task (documents, models, code changes, etc.) | 01-Overview / 01-Introduction |
2.8. Industry-Domain Terms: Creative and Content Industry
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Virtual Try-On | Virtual Try-On | VTON | The image-generation capability of having a target person try on target clothing | 03-Market Research / 03-AI Image group |
| Cross-frame Attention | Cross-frame Attention | — | A mechanism in video generation that models consecutive frames as a whole to suppress character drift | 02-Industry Enablement / 07-Creative Industry group |
| Shot State Machine | Shot State Machine | — | A state machine recording generation state, version, and verification results shot by shot | 02-Industry Enablement / 07-Creative Industry group |
| Brand Asset Library | Brand Asset Library | — | A library carrying cross-project reusable assets such as VI, color palettes, fonts, and character IP | 02-Industry Enablement / 07-Creative Industry group |
| Character Sheet | Character Sheet | — | The anchoring context recording a character’s three views, color palette, and expression set | 02-Industry Enablement / 07-Creative Industry group |
| Explicit Label | Explicit Label | — | A prominent label for AI-generated/synthetic content added at a user-perceivable location | 02-Industry Enablement / 07-Creative Industry group |
| Implicit Label | Implicit Label | — | Attribute information about AI-generated/synthetic content embedded in file metadata | 02-Industry Enablement / 07-Creative Industry group |
| Asset Administration Shell | Asset Administration Shell | AAS | The digital representation of Industry 4.0 components (IEC 63278-1:2023) | 02-Industry Enablement / 07-Creative Industry group |
| Digital Thread | Digital Thread | — | The data chain running through a product’s full lifecycle (ISO 23247-5:2026) | 02-Industry Enablement / 07-Creative Industry group |
| Manufacturing Operations Management | Manufacturing Operations Management | MOM | The L3 boundary layer in the IEC 62264 (ISA-95) five-layer model | 02-Industry Enablement / 07-Creative Industry group |
2.9. Governance, Security, and Compliance
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Role-Based Access Control | Role-Based Access Control | RBAC | A permissions model granting permissions by role; a core component of L6 | 01-Overview / 01-Introduction |
| Proof of Vulnerability | Proof of Vulnerability | PoV | The adjudication method of confirming a vulnerability through a reproducible crash (the core of the AIxCC paradigm) | 02-Industry Enablement / 08-Risk & Compliance group |
| AI Cyber Challenge | AI Cyber Challenge | AIxCC | A DARPA-hosted competition for autonomous vulnerability discovery and patching that established the “model orchestrates, tools judge” paradigm | 02-Industry Enablement / 08-Risk & Compliance group |
| AI Bill of Materials | AI Bill of Materials | AI BOM | The ledger-style management of AI assets within an organization | 02-Industry Enablement / 08-Risk & Compliance group |
| OWASP Top 10 for LLM Applications | OWASP Top 10 for LLM Applications | — | OWASP’s risk list for LLM applications (2025 edition: prompt injection, sensitive-information disclosure, etc.) | 02-Industry Enablement / 08-Risk & Compliance group |
| Excessive Agency | Excessive Agency | — | The risk of granting an agent more permission and autonomy than necessary | 02-Industry Enablement / 03-Software Engineering, 08-Risk & Compliance group |
| Separation of Duties | Separation of Duties | SoD | The governance requirement that generation and review must not share the same identity and credentials (GB/T 45654—2025, clause 4.3.1) | 02-Industry Enablement / 08-Risk & Compliance group |
| PII Scrubbing | PII Scrubbing | — | The identification and scrubbing of personally identifiable information | 02-Industry Enablement / 06-Data Science, 08-Risk & Compliance group |
| Shadow AI | Shadow AI | — | AI tools used privately without governance; a major source of data-leakage incidents | 02-Industry Enablement / 08-Risk & Compliance group |
| Model Risk Management | Model Risk Management | — | The three-pillar development, validation, and governance framework established by SR 11-7 | 02-Industry Enablement / 08-Risk & Compliance group |
| Chain of Evidence / Chain of Custody | Chain of Evidence / Chain of Custody | — | The trace structure allowing conclusions to be traced back to their basis; the main deliverable line of the risk & compliance group | 02-Industry Enablement / 08-Risk & Compliance group |
| Audit Trail | Audit Trail | — | An immutable record of who did what, when, and on what basis | 02-Industry Enablement / 08-Risk & Compliance group |
2.10. Standards Bodies, Regulatory Documents, and Research Institutions
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Agentic AI Foundation | Agentic AI Foundation | AAIF | A foundation under the Linux Foundation (established 2025-12-09) that hosts MCP, goose, and AGENTS.md | 01-Overview / 02-History |
| Series of National Standards for Agent Interconnection | GB/Z 185—2026 | — | China’s series of national standardizing guiding technical documents “Artificial Intelligence – Agent Interconnection,” a closed loop of seven parts, released in the first half of 2026 | 01-Overview / 02-History, 04-Future Outlook |
| Interim Measures for Generative AI Services | Interim Measures for Generative AI Services | — | Order No. 15 of seven ministries’ decree, effective 2023-08-15; Article 19 is the statutory expression of the chain of evidence | 02-Industry Enablement / 08-Risk & Compliance group |
| Measures for Labeling AI-Generated Content | Measures for Labeling AI-Generated Content | — | Effective 2025-09-01, requiring explicit and implicit labels and log retention of no less than six months | 02-Industry Enablement / 07-Creative Industry group |
| Basic Requirements for Generative AI Service Security | GB/T 45654—2025 | — | A national standard implemented 2025-11-01: do not collect when unlawful/bad information exceeds 5%, separate labeling from review personnel, etc. | 02-Industry Enablement / 08-Risk & Compliance group |
| AI Management System | ISO/IEC 42001:2023 | — | The world’s first certifiable AI-management-system standard; Annex A contains 38 controls | 02-Industry Enablement / 03-Software Engineering, 08-Risk & Compliance group |
| AI Risk Management Framework | AI Risk Management Framework | AI RMF | NIST’s voluntary risk-management framework (1.0, 2023-01-26); its GenAI Profile is the generative-AI supplement | 02-Industry Enablement / 08-Risk & Compliance group |
| EU AI Act | EU AI Act, Regulation (EU) 2024/1689 | — | The EU AI regulation in force since 2024-08-01; high-risk obligations include log retention, technical documentation, and human oversight | 02-Industry Enablement / 08-Risk & Compliance group |
| Compliance Management Systems | Compliance Management Systems | — | ISO 37301:2021, identically adopted as GB/T 35770—2022 | 02-Industry Enablement / 08-Risk & Compliance group |
| Model Evaluation & Threat Research | Model Evaluation & Threat Research | METR | The research institution that published a randomized controlled trial of senior developers’ AI efficiency (measuring 19% slower) | 02-Industry Enablement / 03-Software Engineering group |
| DevOps Research and Assessment | DevOps Research and Assessment | DORA | A Google research program publishing annual software-delivery-performance reports | 02-Industry Enablement / 03-Software Engineering group |
| Secure Software Development Framework | Secure Software Development Framework | SSDF | NIST SP 800-218A (v1.1, 2024-10) | 02-Industry Enablement / 03-Software Engineering group |
| Smart Manufacturing Capability Maturity | Smart Manufacturing Capability Maturity | — | The maturity model established by GB/T 39116-2020 | 02-Industry Enablement / 07-Creative Industry group |
2.11. Industry-Domain Terms: AI Infra and Embodied AI
This section was added as the 04-Practice chapter was expanded into eight industry groups, and covers the core terms of the AI Infra group and the Embodied AI group.
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| AI Infrastructure | AI Infrastructure | AI Infra | The stack of compute, frameworks, scheduling, and operations that supports model training and inference; in the six-layer Harness model it corresponds primarily to the physical substrate of the L2 execution layer | 02-Industry Enablement / 01-AI Infra group; 04-Whitepaper / 04-Practice |
| Inference Serving | Inference Serving | — | The engineering step of deploying a trained model as a callable service; core metrics are latency, throughput, and cost per token | 02-Industry Enablement / 01-AI Infra group; 03-Market Research / 06-AI Infra group |
| Training Framework | Training Framework | — | The software stack that organizes the distributed training process, responsible for parallelism strategies, GPU-memory optimization, and fault-tolerant recovery | 02-Industry Enablement / 01-AI Infra group |
| Inference Framework | Inference Framework | — | The deployment-side software stack that optimizes model execution, responsible for operator fusion, quantization, and batch-processing scheduling | 02-Industry Enablement / 01-AI Infra group |
| AI Operations | AI Operations | AIOps | Enhances monitoring, alerting, root-cause localization, and self-healing of clusters and services with AI capabilities | 02-Industry Enablement / 01-AI Infra group |
| Embodied AI | Embodied AI | — | An intelligent system with a physical body that acts in the real environment through a perception–decision–control loop | 02-Industry Enablement / 02-Embodied AI group; 03-Market Research / 07-Embodied AI group |
| Vision-Language-Action Model | Vision-Language-Action Model | VLA | An embodied-AI model paradigm that directly maps visual and linguistic inputs to action outputs | 02-Industry Enablement / 02-Embodied AI group; 04-Whitepaper / 08-Development Outlook |
| Cerebellum (Motor Control Layer) | Cerebellum / Motor Control Layer | — | The layer of an embodied system responsible for high-frequency real-time motion control and posture stability, as opposed to the cerebrum layer responsible for high-level planning | 02-Industry Enablement / 02-Embodied AI group |
| Actuator | Actuator | — | The component that converts control signals into physical motion; the physical endpoint of the L2 execution layer in an embodied system | 02-Industry Enablement / 02-Embodied AI group |
| Sim-to-Real Transfer | Sim-to-Real Transfer | Sim2Real | The method of transferring policies learned in a simulation environment to real physical systems, and the associated gap problem | 02-Industry Enablement / 02-Embodied AI group |
| Compute Scale | Compute Scale | — | The volume of training resources measured by the number of accelerator cards or by compute equivalence; since 2026, one of the competitive barriers for embodied-AI companies | 04-Whitepaper / 08-Development Outlook |
2.12. Version Maintenance and Changelog
This section was added along with the snapshot-level changelog mechanism introduced in the 2026-09-12 snapshot.
| Chinese term | English | Abbr. | Definition | Primary location |
|---|---|---|---|---|
| Snapshot | Snapshot | — | A complete copy of the document library named by date (Harness-Reports-YYMMDD); the day's increments are applied on top of it, and historical snapshots remain read-only | 04-Whitepaper / 11-Incremental Maintenance Whitepaper |
| Incremental Update | Incremental Update | — | An update mode that each day appends only newly added and revised content, without rewriting the entire library | 04-Whitepaper / 11-Incremental Maintenance Whitepaper |
| Manifest | Manifest | — | A machine-readable file recording the digest baseline of every file within a snapshot (MANIFEST.md5), containing full-file MD5s and text fingerprints | Root directory of each snapshot |
| Message-Digest Algorithm 5 | Message-Digest Algorithm 5 | MD5 | The hash algorithm that generates file digests; this project uses its first 8 characters as the human-readable identifier of the change criterion | Root directory of each snapshot |
| Text Fingerprint | Text Fingerprint | — | The content MD5 after removing SVG blocks, figure-caption lines, and source lines, and normalizing consecutive blank lines; used to distinguish ‘text changes’ from ‘figure-only changes’ | Root directory of each snapshot |
| Changelog | Changelog | — | A file-level change description relative to the previous snapshot, listing text changes, figure-only changes, additions, and deletions separately | Root directory of each snapshot |
| Information Cutoff | Information Cutoff | — | The upper time bound of the facts covered by the snapshot's content; items later than that date must not be written into the snapshot | 04-Whitepaper / 11-Incremental Maintenance Whitepaper |
| Evidence Grade | Evidence Grade | A/B/C | A grading of source credibility: A is the official original text, B is authoritative media reporting, C is a general source; items not confirmed by official channels are additionally marked [to be verified] | Whole library |
3. Phrasings Not Adopted by This Whitepaper
The phrasings below have appeared in this project’s corpus or circulate in the industry, but this whitepaper does not treat them as formal terms or arguments. Avoid mixing them in when citing the research library.
- “AI Operating System / Agent OS”: used only as a computer four-layer-stack analogy, not as a formal term. Analogies can aid understanding but cannot shoulder argumentative function.
- “MCP is the USB-C of AI”: appears only in popular-science retellings; this whitepaper’s body does not adopt such analogies as arguments.
- “An authoritative market-size estimate already exists for the Harness layer”: there is no authoritative estimate of the Harness layer’s own market size (
[To be filled]); existing generative-AI market forecasts differ by nearly an order of magnitude, and this whitepaper adopts none of these figures. - “A stronger model will make the Harness disappear”: this whitepaper’s position is that “the Harness will become lighter and be re-architected, but will not disappear,” based on official statements and the phenomenon of post-training coupling.
- “The Agent Framework is the Harness”: not adopted. Frameworks are neutral while a Harness takes a stance; the former is a subset of the latter.
- Anthropomorphic phrasings such as “an agent is a digital employee”: not adopted. This whitepaper uniformly uses “Agent”, the neutral term under China’s official framework.
- “A high leaderboard score equals deployment capability”: not adopted. Public leaderboards can only calibrate the capability ceiling, not predict internal gains (task distributions differ and the metrics are not comparable).
- Unverified vendor-internal figures: e.g. LangChain’s 52.8% vs 66.5% for changing only the Harness, Vercel’s 80% vs 100% for tool pruning, Claude Code’s 350,000 DAU in seven weeks, and Codex Auto-review’s 1/200 and 99% — this whitepaper always retains the label when citing them and does not treat them as established facts.
- Conflating “Agent Harness”, “AI Harness”, and “Agent Runtime”: this whitepaper uniformly uses “AI Harness” for this engineering layer; “Agent Harness” refers only to its implementation form in specific scenarios such as coding; “Agent Runtime” refers only to its runtime product form.
- Unlimited-supply claims of the “LLMs are like water, electricity, and gas” kind: not adopted. Cost, quota, and compute constraints (budget guardrails) are first-class constraints of L6, not unlimited resources.
4. Summary
- This table collects 10 groups of terms, covering core concepts, context and tool protocols, orchestration/memory/evaluation, benchmarks, and the five industry domains of hardware, data, process, creative, and compliance, as well as standards and regulatory institutions, consistent with the terminology of the entire library.
- All definitions follow the original text in the research library; definitions containing numbers inherit the label of the original document.
- The “non-adopted phrasings” listed in Chapter 3 mark the boundary between this whitepaper and some popular online narratives; when citing this whitepaper, defer to this table’s terminology.
- For the full argumentation and context of the terms, see 02-Definitions and the documents of each industry group; for the full canonical-format citation references, see the AGENTS.md / SKILL.md of each group in 04-Practice.
5. Statement of Information Gaps
- The original provenance of the term “Agent Harness”: there is no primary founding literature, so this table does not include a founder entry.
- The official specification texts of newer benchmarks such as SkillsBench and Win-Hill: they appear only in the research library’s retellings; this table’s definitions serve only as an index, and specific evaluation protocols follow the official documents.
- ~~The GB number of the “Measures for Labeling AI-Generated Content”~~: confirmed (original gap closed) — GB 45438—2025, a mandatory national standard, published 2025-02-28 and implemented 2025-09-01, paired with the Labeling Measures (sources: the national-standard full-text public disclosure system std.samr.gov.cn and TC260 official text). This table has incorporated that number.
- The capability-subdomain metric of GB/T 39116-2020: two metrics exist; the research library uniformly adopts PTRM’s four elements / 8 capability domains / 20 capability subdomains and notes the other metric; this table does not elaborate.
- The dedicated meanings of common abbreviations such as DSP and TTM in specific documents: defer to the context of the document in which they appear; this table records only their general definitions.
- Minor metric differences of individual terms across different groups’ documents (e.g. the protocol details of “Handoff (交接)” in multi-agent frameworks): defer to the original text of each direction’s document; this table takes the greatest-common-divisor definition.
6. References
- 01-Overview / 01-Introduction — the source file for unified definitions, the six-layer capability model, and concept boundaries. See `01-概述/01-介绍.md`
- 01-Overview / 02-History — three generations of evolution and key milestones. See `01-概述/02-发展历史.md`
- .workbuddy / PARAM_CARD.md — the single baseline for cross-document consistency. See `.workbuddy/PARAM_CARD.md`
- 02-Industry Enablement / 03-Software Engineering group README — software-engineering terms and measurement metrics. See `02-行业赋能/03-软件工程组/README.md`
- 02-Industry Enablement / 04-Hardware R&D group README — the hardware-domain terminology table. See `02-行业赋能/04-硬件研发组/README.md`
- 02-Industry Enablement / 05-Knowledge Collaboration group README — process and knowledge-domain terms and standard gaps. See `02-行业赋能/05-知识协同组/README.md`
- 02-Industry Enablement / 06-Data Science group README — data and scientific-domain terms. See `02-行业赋能/06-数据科学组/README.md`
- 02-Industry Enablement / 07-Creative Industry group README — content-industry terms and labeling compliance. See `02-行业赋能/07-创意产业组/README.md`
- 02-Industry Enablement / 08-Risk & Compliance group README — governance-and-compliance terms and the regulatory foundation. See `02-行业赋能/08-风险合规组/README.md`
- SWE-bench official site — Princeton NLP, etc. https://www.swebench.com/
- Terminal-Bench official site — Stanford / Laude Institute. https://www.tbench.ai/
- Model Context Protocol official site and spec — MCP / AAIF. https://modelcontextprotocol.io/
- AGENTS.md official site — Linux Foundation / Agentic AI Foundation. https://agents.md/
- Introducing Agent Skills — Anthropic, 2025-10-16. https://www.anthropic.com/news/skills