docs(tutorials): interview-scannability overhaul for #30-33 + collection-wide table-scroll fix (#37)

Responding to owner feedback that the four newest tutorials were
"comprehensive but unscannable before an interview":

- transformer_block: restructured — the concrete Llama-3-like block
  (assembly + end-to-end dataflow + runnable code) moved from a buried
  §7 up to §1, history/design-axis detail now follows ("answer first,
  rationale after"); ~90 internal cross-references renumbered with
  external-tutorial references protected.
- All four de-hedged via a dedicated Codex 5.6-sol scannability review
  (3-way hedge triage: load-bearing/demote/delete) + item-by-item
  rewrites: lead sentence states the conclusion, qualifiers demoted to
  trailing notes, each caveat gets exactly one canonical home.
  inference visible prose -33.5%, transformer_block -17.2%, eval -17.4%,
  pretraining -6.5% bytes. Zero factual changes — the correctness
  boundaries from earlier review rounds are preserved, only relocated.
  Also removed revision-history meta-talk that had leaked into one Q&A
  answer, and 19 "pitfall" sections that merely negated their answers.
- Quantitative hedge-density screen across all 33 tutorials confirmed
  the disease was confined to the four born under the guardrail+review
  process; the two borderline older files (tokenization,
  linear_sparse_attention) were adjudicated healthy by Codex.
- Renderer: wide tables now wrapped in a horizontal-scroll container
  (was clipping); all 67 HTMLs re-rendered to pick it up.
- EN editions fully retranslated post-overhaul (incremental
  section-by-section edits) + fidelity-reviewed: 2 clean PASS, 4 P1s
  fixed in the other two (recurring traps: PRDS expansion, top-p
  "crosses" vs "reaches or exceeds").
- Sidecars refreshed with new source hashes + de-hedge/fidelity rounds;
  strict gate PASS (67 OK / 0 WARN / 0 FAIL).

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: wanshuiyin

README.md

@@ -66,6 +66,7 @@ Phone on the subway, iPad at a café, laptop in the library — same HTML link o
 
 ## 📢 What's New
 
+- **2026-07-31** — ![QUALITY](https://img.shields.io/badge/QUALITY-2ea44f?style=flat-square) ✂️ **Interview-scannability overhaul for #30-33 + collection-wide table-scroll fix** — responding to real reader feedback ("comprehensive but can't grab the key points in a quick pre-interview scan"): **Transformer Block restructured** — the concrete Llama-3-like block assembly + end-to-end dataflow + runnable code moved from a buried §7 up to §1 ("answer first, rationale after"); all four newest tutorials **de-hedged** — lead sentence states the conclusion, qualifiers demoted to trailing notes, each caveat gets exactly one home (Codex 5.6-sol scannability review + item-by-item rewrites; inference visible prose -33.5%, transformer_block -17.2%, eval -17.4%), with zero factual changes; a quantitative screen confirmed the older 29 tutorials are healthy and untouched (tokenization/linear-sparse adjudicated as false positives by Codex). Renderer fixed wide-table overflow (horizontal-scroll wrapper), all 67 HTMLs re-rendered. All four EN editions fully retranslated + fidelity-reviewed (2 clean PASS, 4 P1s fixed across the other 2). Gate PASS.
 - **2026-07-22** — ![NEW](https://img.shields.io/badge/NEW-red?style=flat-square) 🧱 **4 new cheat sheets (#30-33): Transformer Block + LLM Evaluation & Benchmarking + LLM Pretraining Pipeline + LLM Inference & Serving Stack** — filling four previously-missing foundational/systems topics in one batch. **Transformer Block**: unified residual-topology formulas · MHA→MQA→GQA→MLA as independent design axes · Dense FFN vs MoE as an independent capacity axis · the GPT-2→Llama-style evolution, 30 interview questions. **LLM Evaluation & Benchmarking**: pass@k's unbiased combinatorial estimator vs. the biased plug-in (the single easiest place to get this wrong) · evaluator ladder · benchmark contamination detection · Bradley-Terry/Elo, 25 questions. **LLM Pretraining Pipeline**: Kaplan vs Chinchilla scaling-law exponents · corpus-factory data pipeline · document packing/loss masking, 26 questions. **LLM Inference & Serving Stack**: request state machine · exact sampling-operator support-set definitions · PagedAttention/KV lifecycle, 25 questions. All four are the first in this collection to move the cross-model design review before drafting (90-105 numbered guardrails each), then went through 3-5 independent Codex GPT-5.6-sol review batches post-draft — 3 real bugs were each independently caught by two or three separate review batches (Bradley-Terry separation criterion, the best-of-n KV-memory formula). All **bilingual** with runnable scripts and gate PASS.
 - **2026-07-13** — ![QUALITY](https://img.shields.io/badge/QUALITY-2ea44f?style=flat-square) 🔍 **Diffusion-cluster resweep completes the sweep — 8 tutorials, 66 fixes, all 28 tutorials now under GPT-5.6-sol** — covers the diffusion/generative-media cluster deferred from the prior resweep (07-12). Same two-stage pipeline (GPT-5.6-sol finds → independent Claude adversarially verifies): 67 candidates → 66 fixed, 2 REFUTED left untouched; includes 5 real code bugs (iCT/FSQ/LFQ/DDPO) and a recurring "conditional path is straight" ≠ "marginal ODE trajectory is straight" confusion. Gate PASS. ([#32](https://github.com/wanshuiyin/ARIS-in-AI-Offer/pull/32) · [bfae8f1](https://github.com/wanshuiyin/ARIS-in-AI-Offer/commit/bfae8f1))
 - **2026-07-12** — ![QUALITY](https://img.shields.io/badge/QUALITY-2ea44f?style=flat-square) 🔍 **Full-collection resweep — 20 tutorials, 236 fixes, cross-model review upgraded to GPT-5.6-sol** — after the reviewer moved from GPT-5.5 to GPT-5.6-sol, re-audited training fundamentals / attention / RLHF / inference systems / PEFT / agents / RAG·VLM (20 files). Same two-stage pipeline: 240 candidates → 236 fixed, 4 REFUTED left untouched; includes 5 real code bugs and StarPO's acronym settled from the paper's own abstract. Gate PASS. ([#31](https://github.com/wanshuiyin/ARIS-in-AI-Offer/pull/31) · [59636aa](https://github.com/wanshuiyin/ARIS-in-AI-Offer/commit/59636aa))

README_CN.md

@@ -66,6 +66,7 @@
 
 ## 📢 最新动态
 
+- **2026-07-31** — ![QUALITY](https://img.shields.io/badge/QUALITY-2ea44f?style=flat-square) ✂️ **#30-33 四篇"面试可扫读性"专项重构 + 全库表格滚动修复** —— 响应真实读者反馈("内容全但面试前快速看抓不到重点"):**Transformer Block 结构重排**,具体的 Llama-3-like block 组装+端到端数据流+可运行代码从垫底的 §7 提前为 §1,"先给答案、后讲道理";四篇统一**去冗余**——主句先出结论、限定后置成注、同一 caveat 只留唯一归宿(Codex 5.6-sol 可扫读性专项审 + 逐条改写落地,inference 可见正文 -33.5%、transformer_block/-17.2%、eval/-17.4%),事实结论零改动;全库量化筛查确认老 29 篇健康无需动(tokenization/linear-sparse 送 Codex 裁定为假阳性)。渲染器修复宽表格溢出(加横向滚动 wrapper),67 个 HTML 全量重渲染。EN 四篇全量重翻+保真审(2 篇 PASS 零问题,2 篇共 4 处 P1 已修)。Gate PASS。
 - **2026-07-22** — ![NEW](https://img.shields.io/badge/NEW-red?style=flat-square) 🧱 **第 30-33 篇上线：Transformer Block + LLM Evaluation & Benchmarking + LLM Pretraining Pipeline + LLM Inference & Serving Stack** —— 一次性补全四个此前缺失的地基/系统主题。**Transformer Block**：三种残差拓扑统一公式 · MHA→MQA→GQA→MLA 独立设计轴 · Dense FFN vs MoE 容量轴 · GPT-2→Llama-style 演进，30 高频题。**LLM Evaluation & Benchmarking**：pass@k 无偏组合数估计 vs 有偏 plug-in（最易错点）· evaluator ladder · benchmark 污染检测 · Bradley-Terry/Elo，25 高频题。**LLM Pretraining Pipeline**：Kaplan vs Chinchilla scaling law 精确指数 · corpus factory 数据流水线 · document packing/loss masking，26 高频题。**LLM Inference & Serving Stack**：请求状态机 · 采样算子精确支持集定义 · PagedAttention/KV 生命周期，25 高频题。四篇均首次把跨模型设计前置审移到起草之前（每篇 90-105 条护栏），起草后再经 3-5 批 Codex GPT-5.6-sol 独立审核，其中 3 处 bug 被两到三批审核各自独立发现（Bradley-Terry separation 判定、best-of-n KV 公式）。均**双语**+可跑脚本+gate PASS。
 - **2026-07-13** — ![QUALITY](https://img.shields.io/badge/QUALITY-2ea44f?style=flat-square) 🔍 **扩散生成簇重扫收尾 —— 8 篇教程、66 处修复，全集 28 篇正式全部纳入 GPT-5.6-sol** —— 补齐上一轮（07-12）暂缓的扩散/生成媒体簇。同一套两阶段流水线（GPT-5.6-sol 找错 → 独立 Claude 对抗验证）：67 条候选 → 66 处落地，2 条 REFUTED 保留原文；含 iCT / FSQ / LFQ / DDPO 等 5 处真实代码 bug，以及"conditional path 直线"≠"marginal ODE 轨迹直线"这类高频概念混淆。Gate PASS。（[#32](https://github.com/wanshuiyin/ARIS-in-AI-Offer/pull/32) · [bfae8f1](https://github.com/wanshuiyin/ARIS-in-AI-Offer/commit/bfae8f1)）
 - **2026-07-12** — ![QUALITY](https://img.shields.io/badge/QUALITY-2ea44f?style=flat-square) 🔍 **全集重扫 —— 20 篇教程、236 处修复，跨模型审升级到 GPT-5.6-sol** —— 审稿模型从 GPT-5.5 升级后对训练基础 / attention / RLHF / 推理系统 / PEFT / agent / RAG·VLM 共 20 篇重审。同一套两阶段流水线：240 条候选 → 236 处落地，4 条 REFUTED 保留原文；含 5 处真实代码 bug、StarPO 缩写用论文原文定案。Gate PASS。（[#31](https://github.com/wanshuiyin/ARIS-in-AI-Offer/pull/31) · [59636aa](https://github.com/wanshuiyin/ARIS-in-AI-Offer/commit/59636aa)）

docs/index.html

@@ -8,7 +8,7 @@
 <meta name="generator" content="ARIS render-html (academic, v1)">
 <meta name="aris:source-path" content="docs/about.md">
 <meta name="aris:source-sha256" content="15ccd3d65d308c4656117fca714381d163daa0631e2053409704f6c45f4617e8">
-<meta name="aris:generated-at" content="2026-06-16 06:15 UTC">
+<meta name="aris:generated-at" content="2026-07-31 02:58 UTC">
 
 <!-- MathJax 3 -->
 <script>
@@ -238,14 +238,21 @@
 .callout-bad { background: var(--bad-bg); border-left-color: var(--bad); }
 .callout-bad .callout-title { color: var(--bad); }
 
+.table-wrap {
+  overflow-x: auto;
+  margin: 16px 0;
+  border: 1px solid var(--border);
+  border-radius: 4px;
+}
 table {
   width: 100%;
   border-collapse: collapse;
-  margin: 16px 0;
   font-size: 14px;
-  border: 1px solid var(--border);
-  border-radius: 4px;
-  overflow: hidden;
+}
+.table-wrap table {
+  margin: 0;
+  border: none;
+  border-radius: 0;
 }
 thead { background: var(--primary); color: white; }
 th, td {
@@ -546,7 +553,7 @@ <h1>About</h1>
   <div class="meta">
     <span><strong>Source:</strong> <code>docs/about.md</code></span>
     <span><strong>SHA256:</strong> <code>15ccd3d65d30</code></span>
-    <span><strong>Rendered:</strong> 2026-06-16 06:15 UTC</span>
+    <span><strong>Rendered:</strong> 2026-07-31 02:58 UTC</span>
 
   </div>
 </header>
@@ -604,7 +611,7 @@ <h2 id="aris-的-跨模型对抗审查-协议">ARIS 的 "跨模型对抗审查"
 <p>具体案例（来自仓库 audit trail）：</p>
 <ul><li><strong>VLM 教程</strong>：M-RoPE <code>mrope_section=[16,24,24]</code> 单位被纠正（half-dim pair 不是 dim count）；BLIP-2 stage 数 3 → 2；Qwen2-VL vision tower 攻击 DFN-derived 非自训练</li><li><strong>MoE 教程</strong>：Q20 Mixtral 8×7B 算术从「7×8+2.3≈47B」改成 shared 2.4B + 8×4.8B FFN ≈ 40.8B（路由增长到 46.7B）；§4.4 router gradient 补齐 softmax/renorm 完整 Jacobian；8×H100 80G = 640GB ≠ 671B FP8 修正</li><li><strong>Agent Foundations 教程</strong>：跑了 9 轮 codex review，抓到 30+ 个真错（ReAct Fever 数字、SWE-bench Verified OpenAI attribution、MCP 2025-11-25 DCR demoted to MAY、A2A v1.0 SCREAMING_SNAKE_CASE、Anthropic Tool Use GA 2024-05-30 等）</li></ul>
 <h2 id="教程分类">教程分类</h2>
-<table><thead><tr><th>Category</th><th>Topics</th></tr></thead><tbody><tr><td>🧠 General / 基础</td><td>Attention · KL Divergence in RLHF (k1/k2/k3 · placement gradient bias)</td></tr><tr><td>🎯 Post-Training &amp; Reasoning</td><td>RLHF / DPO / GRPO / PPO · Reasoning Models (o1 / R1) · LLM On-Policy Distillation (MiniLLM / GKD / Qwen3 / Tinker)</td></tr><tr><td>🏛️ LLM Architecture &amp; Systems</td><td>MoE (DeepSeek-V3) · Long Context (RoPE / YaRN / MLA) · KV Cache + Speculative Decoding · Quantization (GPTQ / AWQ / FP8 / NVFP4) · Distributed Training (FSDP2 / ZeRO / TP / PP / EP)</td></tr><tr><td>🌊 Generative Models — Theory &amp; Tokenizers</td><td>Flow Matching · Diffusion Foundations · VAE / VQ-VAE / VQ-GAN / FSQ</td></tr><tr><td>🎨 Generation Systems</td><td>Image Gen (SD3 / FLUX / ControlNet) · Video Gen (Sora / Hunyuan-Video / Wan) · 3D Gen (NeRF / 3DGS / SDS) · Diffusion Post-Training (DDPO / DPOK / Diffusion-DPO / Flow-GRPO) · Diffusion / Flow Distillation (CM / sCM / LCM / DMD / DMD2 / ADD)</td></tr><tr><td>👁️ Multimodal</td><td>VLM (CLIP / LLaVA / Qwen-VL)</td></tr><tr><td>🤖 Agents</td><td>Agent Foundations (ReAct / MCP / A2A) · Agentic RL (AgentTuning / ToolRL / RAGEN / WebRL) · Multi-Agent &amp; Long-Horizon (CAMEL / AutoGen / MoA / MemGPT / LATS) · Self-Evolving Agents (Ctx2Skill / Voyager / Reflexion / STaR)</td></tr></tbody></table>
+<div class="table-wrap"><table><thead><tr><th>Category</th><th>Topics</th></tr></thead><tbody><tr><td>🧠 General / 基础</td><td>Attention · KL Divergence in RLHF (k1/k2/k3 · placement gradient bias)</td></tr><tr><td>🎯 Post-Training &amp; Reasoning</td><td>RLHF / DPO / GRPO / PPO · Reasoning Models (o1 / R1) · LLM On-Policy Distillation (MiniLLM / GKD / Qwen3 / Tinker)</td></tr><tr><td>🏛️ LLM Architecture &amp; Systems</td><td>MoE (DeepSeek-V3) · Long Context (RoPE / YaRN / MLA) · KV Cache + Speculative Decoding · Quantization (GPTQ / AWQ / FP8 / NVFP4) · Distributed Training (FSDP2 / ZeRO / TP / PP / EP)</td></tr><tr><td>🌊 Generative Models — Theory &amp; Tokenizers</td><td>Flow Matching · Diffusion Foundations · VAE / VQ-VAE / VQ-GAN / FSQ</td></tr><tr><td>🎨 Generation Systems</td><td>Image Gen (SD3 / FLUX / ControlNet) · Video Gen (Sora / Hunyuan-Video / Wan) · 3D Gen (NeRF / 3DGS / SDS) · Diffusion Post-Training (DDPO / DPOK / Diffusion-DPO / Flow-GRPO) · Diffusion / Flow Distillation (CM / sCM / LCM / DMD / DMD2 / ADD)</td></tr><tr><td>👁️ Multimodal</td><td>VLM (CLIP / LLaVA / Qwen-VL)</td></tr><tr><td>🤖 Agents</td><td>Agent Foundations (ReAct / MCP / A2A) · Agentic RL (AgentTuning / ToolRL / RAGEN / WebRL) · Multi-Agent &amp; Long-Horizon (CAMEL / AutoGen / MoA / MemGPT / LATS) · Self-Evolving Agents (Ctx2Skill / Voyager / Reflexion / STaR)</td></tr></tbody></table></div>
 <blockquote><p>🌐 <strong>双语版本</strong>：23 篇 cheat sheet 每篇都有中文 + 英文 HTML 两个版本（文件名 <code>*_tutorial.html</code> / <code>*_tutorial_en.html</code>）。</p></blockquote>
 <h2 id="怎么贡献">怎么贡献</h2>
 <p>读 <a href="../CONTRIBUTING.md">CONTRIBUTING.md</a>（<a href="../CONTRIBUTING_CN.md">中文</a>）。</p>
@@ -616,7 +623,7 @@ <h2 id="license">License</h2>
   Generated by <a href="https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep/blob/main/skills/render-html/SKILL.md">ARIS <code>/render-html</code></a> ·
   source path <code>docs/about.md</code> ·
   SHA256 <code>15ccd3d65d30</code> ·
-  generated at 2026-06-16 06:15 UTC.
+  generated at 2026-07-31 02:58 UTC.
   This is a generated view — edit the source Markdown, then re-render.
 </footer>
 </main>