# Daily i+1 English Reading - 2026-05-29 # Daily i+1 Reading Recommendations ## Context used - 发现你昨天（2026-05-28）在 `Documents/job-bu/data-analysis-workspace/adhoc/` 集中产出“知识被调用次数分布 / 覆盖率曲线 / Pareto 曲线 / 阈值表”等 CSV+PNG，以及一个把“query recall × knowledge coverage”回填到飞书 Base 的脚本：`2026-05-28_backfill_query_recall_knowledge_to_base.py`（强信号：你在做**知识供给×调用覆盖**的可审计统计闭环）。 - 读取了你昨天的 i+1 清单（2026-05-28），避免重复选同一篇或同一角度（你昨天已读：agent 安全策略 Cedar、eval harness 入门、eval harness 组件、以及一篇 runtime harness 论文）。 - 未能使用：可直接查询的浏览器历史/可读导出数据源（本次未发现现成入口）。 ## Recommendations 1) Inside the LLM Call: GenAI Observability with OpenTelemetry 2. Link: https://opentelemetry.io/blog/2026/genai-observability/ 3. Topic: 用 OpenTelemetry GenAI 语义约定把 agent/LLM/tool 调用“打点成可追踪的分布式系统” 4. Why it matches the user: 你在做“会话→召回→知识→覆盖”的可审计链路；这篇把同样的“链路化思维”落到 **trace/metrics/events**，对你后续做“为什么这次命中了/没命中、慢在哪里、错在哪里”非常对齐 5. Why it is i+1: 句子不难，但会出现一批工程化抽象名词（semantic conventions / span attributes / OTLP 等），非常适合做“可复用写作表达” 6. Estimated new concepts/words/chunks count: 7 7. Likely new concepts or word chunks: - semantic conventions - span attributes / span events - opt in to content capture - structured log events - tool invocation / tool result - token counts / durations - sensitive data redaction (in telemetry) 8. Suggested reading method: 先只读“为什么没 observability 就是在猜”+“默认不采集内容、只采集元数据”两段；然后用你自己的场景写 3 句英文：`We instrument ... so we can attribute failures to ...` 2) Defense in depth for autonomous AI agents 2. Link: https://www.microsoft.com/en-us/security/blog/2026/05/14/defense-in-depth-autonomous-ai-agents/ 3. Topic: 把 agent 安全拆成分层（model/safety/app/positioning），并强调 **application layer = deterministic enforcement** 4. Why it matches the user: 你昨天的工作核心就是“把概率性输出收敛到确定性口径（规则、阈值、覆盖率、回填）”；这篇正好给你一套“可在评审/PRD/复盘里复用”的英文框架与术语 5. Why it is i+1: 逻辑清晰、非学术腔，但大量高频安全/系统设计搭配（bounded scope / least privilege / escalation triggers） 6. Estimated new concepts/words/chunks count: 7 7. Likely new concepts or word chunks: - defense in depth - application layer translates probabilistic → deterministic outcomes - “everything agent” (dangerous failure mode) - bounded scope / scoped permissions - progressive permissioning / start at zero - deterministic human-in-the-loop (HITL) - escalation triggers are defined in code 8. Suggested reading method: 只读“Why the application layer matters”+“Pattern 1/2/3”；每个 Pattern 产出 1 张“英语概念卡”：概念=一句规则 + 你在“知识覆盖回填/阈值表”里的对应例子 3) Fast vs. accurate: Measuring the recall of quantized vector search 2. Link: https://www.elastic.co/search-labs/blog/recall-vector-search-quantization 3. Topic: 用 recall@k + “候选集 oversample→rescore”解释向量检索在速度/精度上的工程折中 4. Why it matches the user: 你昨天在做 Pareto/覆盖率曲线，本质是在做“TopN 覆盖/阈值”；这篇把同样的思路映射到检索系统：先快拿候选，再精排 5. Why it is i+1: 工程叙事短句多、术语不算离谱，但会有一批检索系统常用动词搭配（traverse / rescore / oversample / benchmark） 6. Estimated new concepts/words/chunks count: 6 7. Likely new concepts or word chunks: - recall@10 / recall@k - quantization / compressed vectors - traversal (of the graph) - oversample candidates - rescore with full precision - benchmark against your own data 8. Suggested reading method: 先把“Traversal vs Rescore”抄成两条英文要点，再用你的知识库场景改写：`We first retrieve top-N candidates cheaply, then ...` 4) LLM Readiness Harness: Evaluation, Observability, and CI Gates for LLM/RAG Applications (arXiv) 2. Link: https://arxiv.org/abs/2603.27355 3. Topic: 把评测变成“部署决策工作流”：多指标（groundedness/hit rate/cost/p95 latency）+ Pareto frontiers 4. Why it matches the user: 你昨天已经在用 Pareto/覆盖率做“供给治理决策”；这篇直接提供“把多指标汇总成 readiness、并用 Pareto 避免单指标误导”的英文叙事 5. Why it is i+1: 论文正文更难，但只读 Abstract 就是可控 i+1（学一套研究写作的高复用句式） 6. Estimated new concepts/words/chunks count: 6 7. Likely new concepts or word chunks: - readiness harness - deployment decision workflow - CI quality gates - scenario-weighted scores - retrieval hit rate / groundedness - Pareto frontiers (not a single metric) 8. Suggested reading method: 只读 Abstract；把它改写成你自己的版本（3 句）：`We turn ... into ... by combining ... and ...` ## Vocabulary budget - Estimated daily new-item total: 7 + 7 + 6 + 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；不代表你今天遇到的每个新词/短语都值得做成卡。 ## How to use with Anki - 加到「英语概念卡」：只收“你能在评审/复盘/设计文档里复述并复用”的表达（例如 `deterministic enforcement`, `escalation triggers are defined in code`, `oversample then rescore`, `scenario-weighted readiness`），每张卡必须绑定你自己的例句（来自：知识覆盖率曲线、阈值表、回填脚本的字段与决策口径）。 - 不要加：一次性专有名词堆叠、你不会复用的产品细节、以及你已 mastered/已 suspended 的概念。 - 「阅读词汇量」是 backlog/参考词汇库；「英语概念卡」用于沉淀“带语境、能迁移、能反复用于写作与决策”的概念卡。