deusyu · deusyu · Jul 27, 2026 · Jul 27, 2026 · Jul 27, 2026 · Jul 27, 2026
diff --git a/.claude/skills/curate-research/SKILL.md b/.claude/skills/curate-research/SKILL.md
@@ -1,6 +1,6 @@
 ---
 name: curate-research
-description: 把一批 Harness Engineering 调研候选（文章/论文/工具的 URL）走完「抓取→翻译→评审→收录→清理」流水线，整合进本仓库 works/ 与 references/articles.md，并保持 C1–C12 一致性检查全绿。当用户说"处理这批调研候选 / 收录这些链接 / 整理 translate / 把这几篇翻译收进来"时使用。这是仓库给自己用的策展 harness。
+description: 把一批 Harness Engineering 调研候选（文章/论文/工具的 URL）走完「抓取→翻译→评审→收录→清理」流水线，整合进本仓库 works/ 与 references/articles.md，并保持 C1–C13 一致性检查全绿。当用户说"处理这批调研候选 / 收录这些链接 / 整理 translate / 把这几篇翻译收进来"时使用。这是仓库给自己用的策展 harness。
 ---
 
 # curate-research —— 仓库自我策展 harness
@@ -60,12 +60,14 @@ description: 把一批 Harness Engineering 调研候选（文章/论文/工具
 
 **入库前 checklist（每篇 works/ 候选必过，再进 ⑤）：**
 - [ ] frontmatter 声明 `sourceFigureCount` 并与原文实际图数核对（null = 原文不可得、未审计）；原文插图下载入 `works/imgs/<slug>/` 本地嵌入——C10 校验嵌图数 ≥ 声明数、本地路径文件存在
+- [ ] **数图要抓原文 HTML 数，不能只看渲染出来的正文。** 2026-07-27 的教训：一篇 claude.com 译文声明 `sourceFigureCount: 0`，而原文 HTML 里有 4 个 `<figure>`——抓取工具吐出的 markdown 把它们丢了，人只看那份 markdown 就会以为没有图。可用 `curl <url> | grep -c '<figure'` 交叉验证
+- [ ] 若确实为 0，必须同时写 `sourceFigureAudit`（含 `YYYY-MM-DD`），说明怎么核对的——C13 会拦。判定只算正文配图，站点 logo / 作者头像 / 页脚图标 / 推荐位缩略图 / 社交卡片不计入，但这个判断要写进审计值
 - [ ] 译文保留原文正文超链接（不得译丢）
 - [ ] 关键数字 / 结论句与原文抽查比对（防翻译走样）
 - [ ] 跑 `bash scripts/check-consistency.sh` 全绿
 
 ### ⑤ 校验 + 提交
-- 跑 `bash scripts/check-consistency.sh`，C1–C12 必须全绿再提交。
+- 跑 `bash scripts/check-consistency.sh`，C1–C13 必须全绿再提交。
 - commit 匿名：无 `Co-Authored-By`，作者用 noreply `deusyu@users.noreply.github.com`（依赖全局 git 配置，勿覆盖）。
 - push 仅在用户要求时。
 
@@ -91,4 +93,5 @@ description: 把一批 Harness Engineering 调研候选（文章/论文/工具
 - C8 lint：`source-full.md` 存在时禁止 analysis 谎报摘要页（`scripts/check-consistency.sh`）
 - C9 lint：`concepts/` / `thinking/` / `feedback/` 正文禁止裸写文库计数（历史提法须带"写作时点/当时/此前/首批/首轮/截至/快照"限定词）
 - C10/C11/C12 lint：图片保真（`sourceFigureCount` vs 正文嵌图数 + 本地嵌图路径存在）/ markdown 表格列数与表头一致 / articles.md 编号条目必含 **作者：**/**日期：** 字段（同上脚本）
+- C13 lint：`sourceFigureCount: 0` 必须配 `sourceFigureAudit`（含核对日期）。这条是补 C10 的结构性盲区——C10 的判据是"嵌图数 < 声明数才 FAIL"，所以 0 永远为真，无论你有没有真去数过原文。机械检查只能证伪"多报"，"零报"只能靠留痕让人复核
 - 历史首批实例：articles.md #19–27（2026-06 收录的 9 篇）+ 末尾观察项表 14 行
diff --git a/AGENTS.md b/AGENTS.md
@@ -23,7 +23,7 @@
 - [x] Phase 2：形成自己的观点（thinking/，11 篇，持续中）
 - [x] Phase 3：选一个小项目实践（practice/，1 个 Ralph Demo）
 - [x] Phase 4：记录反馈迭代（feedback/，1 篇，持续中）
-- [x] Phase 5：输出可展示的作品（works/，31 篇翻译 + 1 篇原创 + 2 篇外部中文收录）
+- [x] Phase 5：输出可展示的作品（works/，33 篇翻译 + 1 篇原创 + 2 篇外部中文收录）
 
 > 进度详情以人类向 README.md 的"学习路线"段为准；本节是给智能体的快照。
 
@@ -45,6 +45,10 @@
 - **C7** — 三脉络 per-track 计数（脉络一/二/三）在 4 处下游声明保持一致：READMEs 资料库表、`references/AGENTS.md` 三脉络小标题、`prompts/deep-research-tracker.md` 三脉络明细
 - **C8** — 翻译流水线本地守卫：`translate/<...>/sources/<slug>/source-full.md` 存在时，对应 `01-analysis.md` 不得再声称"仅摘要页 / 建议补抓全文"。`translate/` 已 gitignore，CI 与干净 clone 自动 SKIP，仅本地有过程稿时触发
 - **C9** — `concepts/` / `thinking/` / `feedback/` 正文不得裸写文库计数（"N 篇文章 / N 篇翻译 / N 大概念"）；历史性提法须带"写作时点 / 当时 / 此前 / 首批 / 首轮 / 截至 / 快照"限定词，否则去数字改链 `references/articles.md`
+- **C10** — 图片保真（纯本地、零网络）：每篇 `works/*-translation.md` 的 frontmatter 必须声明 `sourceFigureCount`（缺失即 FAIL；`null` = 原文不可得、未审计 → SKIP；数字 N → 正文嵌图数须 ≥ N），且所有本地嵌图路径（`imgs/...`）必须在磁盘上存在
+- **C11** — markdown 表格形状：README ×2、`references/AGENTS.md`、`references/articles.md`、`works/AGENTS.md` 里每一行表格的单元格数须与表头一致
+- **C12** — 条目字段完整性：`references/articles.md` 每个 `### N.` 编号条目必须带 **作者：** 与 **日期：** 字段
+- **C13** — 零插图声明须留痕：C10 只能证伪"多报"（嵌图数 < 声明数才 FAIL），因此 `sourceFigureCount: 0` 在本地**永远无法被证伪**——不管你有没有真去核对原文，它都是绿的。2026-07-27 就是这个洞放行了一个假 0（原文实有 4 张配图）。C10 刻意零网络、无法回查原文，所以改为要求留痕：**声明 0 的译文必须同时带 `sourceFigureAudit` 字段，值里要有 `YYYY-MM-DD` 核对日期**，写清怎么核对的、结论是什么。`null` 仍然 SKIP——它本来就自陈未审计
 
 执行：`bash scripts/check-consistency.sh`（仓库根目录）
 启用 pre-commit 阻断：`git config core.hooksPath .githooks`

diff --git a/README.en.md b/README.en.md
@@ -1,8 +1,8 @@
 [中文](README.md) | English
 
 ![License: MIT](https://img.shields.io/badge/license-MIT-blue)
-![Articles](https://img.shields.io/badge/articles-61-green)
-![Translations](https://img.shields.io/badge/translations-31-orange)
+![Articles](https://img.shields.io/badge/articles-73-green)
+![Translations](https://img.shields.io/badge/translations-33-orange)
 
 # Harness Engineering Study Guide
 
@@ -112,10 +112,10 @@ harness-engineering/
 ├── thinking/              # Phase 2: Independent analysis (11 articles)
 ├── practice/              # Phase 3: Hands-on experiments (1 Ralph Demo)
 ├── feedback/              # Phase 4: Lessons learned (1 article)
-├── works/                 # Phase 5: Shareable outputs (31 translations + 1 original + 2 external Chinese captures)
+├── works/                 # Phase 5: Shareable outputs (33 translations + 1 original + 2 external Chinese captures)
 ├── tools/                 # Tools that reduce the 6 complexity dimensions
 ├── prompts/               # Validated prompts collection
-└── references/            # External resource index (61 articles with deep summaries)
+└── references/            # External resource index (73 articles with deep summaries)
 ```
 
 Each subdirectory has its own `AGENTS.md` explaining its purpose and conventions — a direct practice of the "progressive disclosure" principle from the original article.
@@ -126,15 +126,15 @@ Each subdirectory has its own `AGENTS.md` explaining its purpose and conventions
 - [x] **Phase 2: Form your own opinions** — 11 independent analyses (ongoing)
 - [x] **Phase 3: Pick a small project to practice** — Ralph Demo completed (321s, $0.31)
 - [x] **Phase 4: Record feedback & iterations** — 1 article (ongoing)
-- [x] **Phase 5: Produce shareable work** — 31 professional translations + 1 original synthesis + 2 external Chinese captures
+- [x] **Phase 5: Produce shareable work** — 33 professional translations + 1 original synthesis + 2 external Chinese captures
 
 ## 📚 Research Library
 
-61 articles across three knowledge tracks + 2 extended readings:
+73 articles across three knowledge tracks + 2 extended readings:
 
 | Track | Coverage | Perspectives |
 |-------|----------|-------------|
-| AI-Era Harness Engineering | 57 articles | OpenAI → Fowler → Anthropic → LangChain → Stanford → Claude Code reverse engineering & source leak → Subagent runtime → Sensors/SPDD/ADLC → Out-of-scope, safety auditing & quality postmortems → Evaluation trilogy → Dynamic workflows → Origins (Ralph / Hashimoto) & discipline synthesis → Codex harness anatomy → Loop Engineering trilogy → Self-evolving harnesses & RSI → Formal verification → Multi-agent scaling (Cursor / C compiler) → Official containment & evals methodology → Behavior maps / DSLs / local models / outer-loop accountability → industrial-scale mechanical porting (Bun) & harness-model co-evolution (HarnessX) |
+| AI-Era Harness Engineering | 69 articles | OpenAI → Fowler → Anthropic → LangChain → Stanford → Claude Code reverse engineering & source leak → Subagent runtime → Sensors/SPDD/ADLC → Out-of-scope, safety auditing & quality postmortems → Evaluation trilogy → Dynamic workflows → Origins (Ralph / Hashimoto) & discipline synthesis → Codex harness anatomy → Loop Engineering trilogy → Self-evolving harnesses & RSI → Formal verification → Multi-agent scaling (Cursor / C compiler) → Official containment & evals methodology → Behavior maps / DSLs / local models / outer-loop accountability → industrial-scale mechanical porting (Bun) & harness-model co-evolution (HarnessX) → long-running harness foundations & eval-environment confounders (Anthropic backfill) → harness operations metrics & reward hacking (Cursor backfill) → tool schemas are not neutral → the software-factory debate (Dex Horthy / Osmani) → agent-swarm cost economics → deleting 80% of the system prompt |
 | Cloud-Native Harness.io | 2 articles | CI/CD platform architecture (same name, different meaning) |
 | Efficiency Paradox & Capability Evolution | 2 articles | YDD systematic teardown + METR follow-up (measurement-methodology crisis) |
 | Extended Reading | 2 articles | Context Engineering, Human-Agent collaboration |
@@ -144,11 +144,13 @@ See [references/articles.md](references/articles.md) — each article includes c
 ## 📖 Translations
 
 <details>
-<summary><b>31 Chinese translations of key articles</b> (click to expand)</summary>
+<summary><b>33 Chinese translations of key articles</b> (click to expand)</summary>
 
 | Translation | Original Author | Source |
 |-------------|----------------|--------|
 | ⭐ [Eight Years of Wanting](works/maganti-eight-years-building-ai-translation.md) | Lalit Maganti | Personal blog |
+| [The New Rules of Context Engineering for Claude 5](works/anthropic-context-engineering-claude5-translation.md) | Thariq Shihipar | Anthropic / Claude |
+| [Better Models: Worse Tools](works/ronacher-better-models-worse-tools-translation.md) | Armin Ronacher | Personal blog |
 | [Rewriting Bun in Rust](works/bun-in-rust-translation.md) | Jarred Sumner | Bun Blog |
 | [Building a C Compiler with a Team of Parallel Claudes](works/anthropic-c-compiler-translation.md) | Nicholas Carlini | Anthropic |
 | [Scaling Long-Running Autonomous Coding](works/cursor-scaling-agents-translation.md) | Wilson Lin | Cursor |
@@ -220,7 +222,7 @@ The "Ralph Wiggum Loop" is the core implementation pattern of Harness Engineerin
 
 ## 🛠️ Development Notes
 
-The repo ships with a consistency checker, `scripts/check-consistency.sh`, guarding against count drift across nine layers of checks:
+The repo ships with a consistency checker, `scripts/check-consistency.sh`, guarding against count and fidelity drift across thirteen layers of checks:
 
 - **C1-C2** — `references/articles.md` article count + its 4 downstream claim sites (README × 2 badges, `prompts/deep-research-tracker.md` header, `references/AGENTS.md` overview)
 - **C3** — actual `*.md` file counts in `concepts/` / `thinking/` / `feedback/` match the README "X 篇" claims
@@ -230,6 +232,10 @@ The repo ships with a consistency checker, `scripts/check-consistency.sh`, guard
 - **C7** — per-track counts (Track 1/2/3) stay consistent across their 4 downstream claim sites (README research-library tables × 2, `references/AGENTS.md` track headings, `prompts/deep-research-tracker.md` track lines)
 - **C8** — local translation-pipeline guard: once `translate/<...>/sources/<slug>/source-full.md` is captured, the matching `01-analysis.md` may no longer claim "abstract-only / fetch full text later". `translate/` is gitignored, so this auto-SKIPs on CI and clean clones
 - **C9** — authored prose in `concepts/` / `thinking/` / `feedback/` must not restate library counts ("N articles / N translations") as live facts; historical mentions must carry a dated-snapshot qualifier, otherwise drop the number and link `references/articles.md`
+- **C10** — figure fidelity (purely local, zero network): every translation's frontmatter must declare `sourceFigureCount`, the body must embed at least that many images, and every local embed path must exist on disk (`null` = source unavailable / unaudited → SKIP)
+- **C11** — markdown table shape: in the checked files, every table row must carry the same cell count as its header
+- **C12** — every numbered entry in `references/articles.md` must carry the **作者：** and **日期：** fields
+- **C13** — zero-figure claims need an audit trail. C10 can only falsify OVER-claiming, so `sourceFigureCount: 0` is unfalsifiable locally — that hole shipped a false 0 on 2026-07-27 (the source had 4 body figures). Any translation claiming 0 must therefore also carry `sourceFigureAudit` containing a `YYYY-MM-DD` date, stating how the claim was verified
 
 **Enable the pre-commit hook after first clone:**
 
@@ -249,7 +255,7 @@ See the "机械化检查" section of the root `AGENTS.md` for details.
 
 > This archive now curates itself.
 >
-> Bringing in outside research no longer runs on vibes — it follows a pipeline frozen into a skill, [`curate-research`](.claude/skills/curate-research/SKILL.md): review is automated by parallel agents (the feedback loop), `scripts/check-consistency.sh` keeps counts from drifting via C1–C9 (the mechanical rail), and whether something gets in is always a human gate (humans steer, agents execute).
+> Bringing in outside research no longer runs on vibes — it follows a pipeline frozen into a skill, [`curate-research`](.claude/skills/curate-research/SKILL.md): review is automated by parallel agents (the feedback loop), `scripts/check-consistency.sh` keeps counts and fidelity from drifting via C1–C13 (the mechanical rail), and whether something gets in is always a human gate (humans steer, agents execute).
 >
 > So the constraints themselves became the product — exactly what [concepts/07-spec-as-product.md](concepts/07-spec-as-product.md) argues, except this time the subject is the repo itself.