# Daily i+1 English Reading - 2026-05-26 # Daily i+1 Reading Recommendations ## Context used - 读取了昨日（2026-05-25）的日报：主线仍是 **agentic coding 产品化**、**MCP/tool-use 的安全与审计**、以及 **RAG/检索评测口径（groundedness/faithfulness + retrieval metrics）**。 - 扫描了昨日（2026-05-25）本地修改文件：你在推进两条线——`Documents/product-bu/20-products/tense-hud/...`（浏览器扩展/可视化规则/测试）与 `Documents/learning-bu/english/04-projects/英文歌项目/a-thousand-years/...`（Anki 歌词卡流水线与导入/备份）。 - 未能使用：可直接检索的浏览器历史/可读导出数据源（本次未发现现成入口）。 ## Recommendations 1) Model Context Protocol (MCP): Security Design Considerations for AI-Driven Automation 2. Link: https://www.nsa.gov/Portals/75/documents/Cybersecurity/CSI_MCP_SECURITY.pdf?ver=bmgiSbNQLP6Z_GiWtRt6bg%3D%3D 3. Topic: MCP 在真实落地中暴露的安全缺口（访问控制、审批、token/session、序列化/反序列化、信任链与任务传播）与可执行的设计建议 4. Why it matches the user: 你正在做“多工具编排 + 审计字段 + approvals”，这份是**直接可抄成你产品安全基线**的官方语言 5. Why it is i+1: 英文偏“安全架构报告体”，句式可读但抽象名词密集，适合沉淀成可复用 chunk（而不是背细节） 6. Estimated new concepts/words/chunks count: 8 7. Likely new concepts or word chunks: - underspecified design - interaction pattern inversion - attack path / untraceable access - unverified task propagation - role-based access control (RBAC) - informed consent / approval workflow - bearer token lifecycle (rotation / revocation) - idempotency / message replay 8. Suggested reading method: 只读 Executive summary + “Security concerns”各小节，每小节用一句话写成你系统里的**控制点描述**（英文），并补一个“bad case / good case”对照例句 2) AI Agents May Always Fall for Prompt Injections 2. Link: https://arxiv.org/abs/2605.17634 3. Topic: 把 prompt injection 从“数据-指令分离”防御范式的局限，提升到“上下文完整性（Contextual Integrity）/信息流规范”的框架 4. Why it matches the user: 你关心的是**工具调用边界、权限边界、审计可解释**；这篇给你一个更“原则化”的威胁建模语言 5. Why it is i+1: 论文写作体比新闻难一点，但只读 Abstract + Introduction + Threat framing 就足够，且术语高复用 6. Estimated new concepts/words/chunks count: 6 7. Likely new concepts or word chunks: - prevailing defense paradigm - data–instruction separation - contextual manipulation - contextual integrity (CI) - information flow / norms - impossibility result / adversary can always construct… 8. Suggested reading method: 目标不是通读；抓住“defense fails because…”的因果句式，改写成你自己的场景：**浏览器扩展/工具输出/检索内容**如何跨越权限边界 3) What is RAG evaluation? Measuring retrieval quality and answer groundedness 2. Link: https://www.braintrust.dev/articles/what-is-rag-evaluation 3. Topic: RAG 评测拆解为 retrieval stage 与 generation stage；把“问题在哪”从模糊信号变成可诊断指标组合 4. Why it matches the user: 你近期就在统一 groundedness/faithfulness 与检索指标口径；这篇适合当你评测面板的“英文定义底稿” 5. Why it is i+1: 难点集中在“度量/对比/定义”表达（precision/recall/ranking/traceable claims），但整体对 TOEFL 90 友好 6. Estimated new concepts/words/chunks count: 7 7. Likely new concepts or word chunks: - failure mode - actionable diagnosis - context precision / context recall - retrieval ranking metrics (NDCG@K, Precision@K) - trace a claim back to… - partially grounded / partially hallucinated - golden dataset / regression test 8. Suggested reading method: 读完后立刻产出一张你自己的两列表（Retrieval / Generation），每个指标写：定义 + 你系统里对应的 log/trace 字段名 4) Optimize long tasks 2. Link: https://web.dev/articles/optimize-long-tasks 3. Topic: 主线程 long task 的成因与拆分策略（把 >50ms 的任务切片、调度到合适时机/线程），面向 DevTools 诊断与工程改造 4. Why it matches the user: 你昨天在改浏览器扩展（`tense-hud`）并有测试/可视化预览；这篇能直接转成“扩展卡顿治理清单” 5. Why it is i+1: 文档结构清晰，i+1 来自性能术语与动词搭配（break up / yield / schedule / offload） 6. Estimated new concepts/words/chunks count: 6 7. Likely new concepts or word chunks: - monopolize the main thread - break up work into chunks - yield back to the main thread - schedule work (at the right place/time) - offload work (off the main thread) - responsiveness / jank 8. Suggested reading method: 只读“识别 long tasks + 典型优化手段”两段，然后把你的扩展里一段重逻辑用一句英文描述成：**We break X into Y to avoid long tasks.** ## Vocabulary budget - Estimated daily new-item total: 8 + 6 + 7 + 6 = 27（≥20） - Back-calculate: `14678 / 27 ≈ 544` 天，约 `544 / 365 ≈ 1.49` 年 - 说明：这是“规划预算”，不是承诺；只有高复用、能写出你自己的语境例句、能迁移到产品/文档表达里的项才值得做卡 ## How to use with Anki - 加到「英语概念卡」：优先收“可复用 chunk + 你自己的工程语境例句”，例如 `approval workflow / informed consent`、`unverified task propagation`、`trace a claim back to…`、`monopolize the main thread`（都要绑定你自己的工具/扩展/评测面板场景）。 - 不要加：一次性新闻事实、你不会在写作/PRD/审计里复用的名词堆叠、以及你已 mastered/已 suspended 的概念。 - 「阅读词汇量」是 backlog/参考词汇库；真正需要你“带着语境能复述、能落字段/落文档”的内容，才进入「英语概念卡」。