# Daily i+1 English Reading - 2026-05-25
# Daily i+1 Reading Recommendations
## Context used
- 读取了昨日(2026-05-24)的日报:你近期主线是 **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)**。
- 扫描了近 48 小时本地内容:`odyssey/0 收集箱/Agent健身房/2026-05-24/早间复盘.md` 明确在补“浏览器扩展性能诊断器、helper/watchdog、Surge 预检、roster provenance”等“反复摩擦点”工具化。
- 未能使用:可直接检索的浏览器历史/导出数据源(本次未发现现成可读入口)。
## Recommendations
1) US government, allies publish guidance on how to safely deploy AI agents
2. Link: https://cyberscoop.com/cisa-nsa-five-eyes-guidance-secure-deployment-ai-agents/
3. Topic: Five Eyes 对“安全部署 AI agents”的联合指导要点与风险类别
4. Why it matches the user: 你在做“多工具编排 + 人类决策点 + 审计”的控制平面,这篇能把“该做哪些控制”压缩成可落地清单语言
5. Why it is i+1: 新闻体总体可读,但安全与治理类抽象名词密集,适合沉淀成概念卡(而不是记事实)
6. Estimated new concepts/words/chunks count: 6
7. Likely new concepts or word chunks:
- secure deployment
- guardrails
- threat model / adversary model
- data exfiltration
- privilege escalation
- governance / oversight
8. Suggested reading method: 只读“指导发布的动机 + 风险分类 + 建议控制点”,每类风险写 1 条你产品里的对应控制句(英文),用于 README/审计文档复用
2) Tool Annotations as Risk Vocabulary: What Hints Can and Can't Do
2. Link: https://blog.modelcontextprotocol.io/posts/2026-03-16-tool-annotations/
3. Topic: MCP 工具注解/提示(hints/annotations)怎样变成“风险词汇表”,以及它们能/不能解决什么
4. Why it matches the user: 你正在把工具能力、审批文案、审计字段体系化;这篇直接提供“把风险语义前置到工具层”的表达框架
5. Why it is i+1: 工程随笔可读,但抽象对比句式多(can/can’t, should/must, enforce/communicate),正好练表达
6. Estimated new concepts/words/chunks count: 7
7. Likely new concepts or word chunks:
- risk vocabulary
- capability hint / annotation
- untrusted content
- least privilege / principle of least privilege
- approval prompt / consent prompt
- attack surface
- enforcement vs. signaling
8. Suggested reading method: 每段只抓 1 个“能直接映射到你系统字段/按钮”的 chunk(例如 approval prompt),立刻写一条你自己的例句(含 kanbots/工具审批场景)
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:retrieval-side 指标 + answer-side 指标(groundedness/faithfulness 等)如何组织成可执行 rubric
4. Why it matches the user: 你昨天就围绕“检索质量 vs 回答可信”在做口径统一;这篇适合拿来当你评测面板的术语底稿
5. Why it is i+1: 专业科普型,TOEFL 90 基本能跟上;新点集中在“评测写作句式 + 指标命名”
6. Estimated new concepts/words/chunks count: 7
7. Likely new concepts or word chunks:
- groundedness / grounded in the provided context
- faithfulness vs. relevance
- context relevance
- scoring rubric / evaluation rubric
- LLM judge / evaluator
- retrieval quality
- end-to-end evaluation
8. Suggested reading method: 读完立刻做一张两列小表(Retrieval-side / Answer-side),每个指标写“定义 + 你系统里对应可观测信号(log/trace)”
4) Analyze runtime performance (Chrome DevTools)
2. Link: https://developer.chrome.com/docs/devtools/performance/
3. Topic: 用 DevTools 的 Performance 能力做运行时性能分析(定位长任务、主线程阻塞、CPU 时间等)
4. Why it matches the user: 你在做“扩展卡死性能诊断器”的只读定位链路;DevTools 的术语与操作路径能直接对齐你的诊断输出
5. Why it is i+1: 文档结构清晰、可跳读;但性能分析术语(long task、main thread、flame chart)会带来稳定增量
6. Estimated new concepts/words/chunks count: 6
7. Likely new concepts or word chunks:
- runtime performance
- main thread
- long task
- flame chart
- call stack
- CPU profiling / profiling overhead
8. Suggested reading method: 不要通读;只读“如何录制一次 trace + 如何从 long task 反推原因”两段,然后对照你的诊断器设计补 3 个输出字段名(英文)
## Vocabulary budget
- Estimated daily new-item total: 6 + 7 + 7 + 6 = 26(≥20)
- Back-calculate: `14678 / 26 ≈ 565` 天,约 `565 / 365 ≈ 1.55` 年
- 说明:这是“规划预算”,不是承诺;只有高复用、能写出你自己的正反例句、能迁移到工作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 + 你自己的系统语境例句”,例如 `approval prompt`、`least privilege`、`grounded in context`、`long task`,并写清你产品/工具里对应的控制点或诊断信号
- 不要加:一次性新闻事实、人名机构名、你无法举例的纯术语堆叠、以及已 mastered/已 suspended 的概念
- 「阅读词汇量」是你的 backlog/参考库;真正用于思考与写作表达的“语境化概念”,才进入「英语概念卡」