refactor(book): unify the two competing security taxonomies into one three-layer skeleton - #860
Merged
Conversation
The Chinese source had reader-visible figure numbering gaps left behind by earlier content removals, and the caption number no longer matched the filename in two chapters. - ch5: restore the 图5-4 embed dropped by d39b1d7 while its prose reference survived, closing the 5-3 -> 5-5 gap. - ch7: retire the two figures the chapter dropped to descriptive names and shift fig7-18/19/20 down so caption number == filename again. - ch10: renumber 10-3..10-13 to 10-2..10-11 across all 13 editions, closing the gaps at 10-2 and 10-10; retire the two removed figures to descriptive names. Drop the dangling "Figure 10-10" reference in the English edition. - Rebuild fig8-1 for the 11 editions still shipping the pre-04c0ed70 artwork. The Chinese source now has gapless, filename-aligned numbering in every chapter with no dangling cross-references. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Commit 04c0ed7 replaced all five Chinese chapter-8 figures when the chapter was reorganized around continual evolution. Only book-ko, added afterwards, picked them up; every other edition still shipped the pre-reorganization artwork, so the captions described the new chapter while the pictures showed the old one. In book-en, book-ar and book-zhtw fig8-3 was additionally still untranslated Simplified Chinese, and book-es/ja/tr carried a further two-figure offset whose fig8-3/fig8-4 showed chapter-4 material. Rebuilds fig8-1..fig8-5 for all 11 affected editions from the current Chinese geometry, translating text nodes only so the drawings stay identical, and drops the orphaned fig8-6/fig8-7 left in book-es, book-ja and book-tr. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Eight editions still carried chapter-3 figures from an earlier revision, two
of which showed the wrong drawing entirely:
- fig3-2 ("four memory strategies") rendered the RAG query flow, i.e. the
artwork belonging to fig3-5.
- fig3-4 ("multi-type memory architecture") rendered the HNSW index structure
from fig3-7.
- fig3-1 showed the retired RAG-fundamentals map instead of the chapter
knowledge map (fixed for book-en in #816, never carried across).
- fig3-11 showed the retired GraphRAG walkthrough built on the x86 SSE/AVX
instruction set -- a CPU diagram in a knowledge-graph section. In book-hu it
instead showed fig3-14's contextual retrieval.
- fig3-12 was an earlier revision of the agentic vs. non-agentic comparison.
All five are regenerated for ar, es, hu, id, ja, ru, ta and tr from the current
book-en geometry, translating text nodes only, so every edition now matches the
Chinese drawing exactly.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every edition captioned fig9-7 as "9-6", duplicating the number already used by fig9-6, so all later captions ran one behind their file and the chapter ended with fig9-12 labelled 9-11. Captions are now renumbered to match their filenames, and the fig9-11/fig9-12 embeds (VLA, Sim2Real) are dropped to match the Chinese chapter, which removed them in #730. The files stay on disk; the course slides link fig9-11 directly. Also rebuilds fig4-9 from the Chinese for the six editions that shipped it untranslated or, in book-es and book-tr, showing a duplicate of fig4-8, and switches book-hu to the shared fig2-7.png attention heat map that the other twelve editions use. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every edition captioned fig9-7 as "9-6", duplicating the number already used by fig9-6, so each later caption ran one behind its file and the chapter ended with fig9-12 labelled 9-11. Captions now match their filenames in all twelve editions; the figure set itself is unchanged, so the prose cross-references still resolve. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The chapter 10 renumber matched captions on the "Рис." prefix and missed the one caption written as "Рисунок 10-11", leaving it pointing at fig10-9. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ering Chapter 9's numbering gaps at 9-5/9-6 were a prose problem, not a figure problem: eight editions had condensed the "cognitive timing" section into a single summary paragraph, so the fast/slow-thinking and Step-Audio R1 figures had nowhere to attach. Their SVGs were already present and correctly translated -- just never referenced. - Restore the three #### subsections the Chinese has (fast thinking answers / fast thinking interacts / end-to-end unification) in ar, es, hu, id, ru, ta, tr and vi, translated to match each edition's existing terminology, and re-attach fig9-5 and fig9-6. - Drop the fig9-11/fig9-12 embeds and their in-text references in all twelve editions: #730 deliberately removed both from the Chinese chapter, and the Chinese is authoritative. The VLA and Sim2Real prose stays. - book-hu: move the fig7-7 experiment box back ahead of the pre-training section and the fig10-9 MetaGPT section after fig10-8, matching the Chinese reading order. Chapter 9 is now exactly ten figures in all thirteen editions, and every chapter of every edition is gapless, caption-aligned and in order. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…w, ja, es, ru, tr) Commit 05ab7f3 condensed the Chinese reward-design section into four subsections; the translations still carried the superseded version, roughly ten times longer and organised differently, along with fig7-16 and fig7-17 which the Chinese no longer has. Replaces the three old reward subsections with a translation of the current Chinese four-subsection structure and adds the [^ch7-23] and [^ch7-29] footnotes, which existed only in the Chinese. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Commit 05ab7f3 condensed the Chinese reward-design section from three loosely organised subsections into four sharper ones and dropped two figures. The translations never followed, so every edition still shipped the superseded version -- roughly ten times longer than the Chinese and structured differently -- along with fig7-16 and fig7-17, which no longer exist upstream. - Replace the old reward subsections in ar, en, es, hu, id, ja, ko, ru, ta, tr, vi and zhtw with a translation of the current Chinese four-subsection structure, matching each edition's established terminology. - Add the [^ch7-23] and [^ch7-29] footnotes, previously Chinese-only. - Retire fig7-16/fig7-17 to descriptive names and shift fig7-18/19/20 down, so chapter 7 is 18 figures with caption number == filename in every edition, exactly as the Chinese source now is. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
fig4-7 and fig4-8 shipped with untranslated Chinese labels in book-en, book-ar, book-ta and book-vi; the Arabic copies were additionally part-machine-translated into garbled mixtures. Both are rebuilt from the Chinese geometry with proper translations, keeping tool identifiers and protocol role names as-is. Also translates the one English sentence left in book-ta's fig9-4. The Chinese text remaining in fig2-6 and fig10-5 is deliberate: it is the example sentence being analysed and the glossary data the figure illustrates. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…n/numbering-alignment # Conflicts: # book-ar/chapter7.ar.md # book-ar/chapter9.ar.md # book-en/chapter7.md # book-en/chapter9.md # book-es/chapter7.es.md # book-es/chapter9.es.md # book-hu/chapter10.md # book-hu/chapter7.md # book-hu/chapter9.md # book-id/chapter7.md # book-id/chapter9.md # book-ja/chapter7.ja.md # book-ja/chapter9.ja.md # book-ko/chapter7.ko.md # book-ko/chapter9.ko.md # book-ru/chapter7.md # book-ru/chapter9.md # book-ta/chapter7.ta.md # book-ta/chapter9.ta.md # book-tr/chapter7.tr.md # book-tr/chapter9.tr.md # book-vi/chapter7.vi.md # book-vi/chapter9.vi.md # book-zhtw/chapter7.zhtw.md # book-zhtw/chapter9.zhtw.md # book/chapter10.md # book/chapter7.md # book/images/fig8-17.svg # book/images/fig8-18.svg # slides/public/images/fig10-13.svg
…nslations Bring the numbering skeleton of every translated edition into line with the Chinese edition, so that later structural work can be applied mechanically. The Chinese edition is the source of truth throughout. Experiments - ch2: add the missing 实验 2-7 (writing Skill from personal samples) to all 12 editions and shift the following experiments 2-7..2-9 → 2-8..2-10; rename the accompanying footnote ch2-7 → ch2-8 to match. - ch4: add the missing multimodal-extraction experiment as 4-2 to 10 editions and shift 4-2..4-6 → 4-3..4-7. book-en had this experiment mislabelled as "Experiment 3-7" (colliding with a real 3-7 in chapter 3) and placed before 4-1; it is relabelled 4-2 and moved after the 4-1 box. Chapter summaries that enumerate the experiments are updated (six → seven, ranges corrected). - ch5: book-hu had the production-log experiment mislabelled 5-5 (duplicating the real 5-5) and was missing 5-7 entirely; the box is renumbered 5-8 and a translation of 实验 5-7 (adaptive log parser) is added. - ch7: add the missing 7-12 (V-IRL-VL) and 7-13 (SimpleVLA-RL) boxes to all 12 editions, together with the [^ch7-24] footnote they cite. Also fix the stale in-text numbers: ReTool was cited as 7-15 (it is 7-14) and RLVP as 7-14 (it is 7-16). Figures - ch9: 8 editions compressed the fast/slow-thinking section into a single paragraph and therefore lost 图9-5 and 图9-6. The three-solution block is translated in full so all editions carry the same figure inventory. Tables - ch6: remove the translation-only Pass@k/Pass^k table (no counterpart in the Chinese) and renumber 6-4..6-6 → 6-3..6-5. In book-en this table also duplicated the number of the memory-system table. - ch10: remove the translation-only shared/non-shared selection-criteria table and renumber 10-2..10-4 → 10-1..10-3. - ch9: add the 表9-1 caption to the MiniCPM-o experiment in 11 editions. - book-id ch3: translate the untranslated "Table 3-3" caption. - book-ar ch7: fix the transposed table number 1-7 → 7-1. Other consistency fixes - book-hu: 542 headings and labels across all chapters had their bold markers replaced by straight double quotes, which hid whole experiment boxes from the numbering. Restored (code blocks left untouched; the edition uses „…" for real quotations). - book-hu ch10: move the 10-4 experiment box back into the manager-pattern section so 图10-8 precedes 图10-9. - book-ar ch2: drop an Arabic-only passage (and its [^ch2-5] footnote) that has no counterpart in the Chinese. - book-ru ch2: add the missing [^lost-in-the-middle] footnote. - book-vi ch1: Thử nghiệm 1.2 → 1-2. After this change the only remaining gap is book-ar/chapter4, which is covered by the separate chapter-4 retranslation PR. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Picks up the Arabic chapter 4 retranslation (#857), which closes the one gap the numbering audit had left open. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…2 translations Mirror the Chinese restructure (#856) in every translated edition, so that all 13 editions share one chapter layout, numbering scheme and file naming. Structure - New Chapter 6 "Interaction: Expanding the Observation and Action Spaces", assembled from a newly written framing section, the event-driven async block lifted out of Chapter 4, and the whole of the old Chapter 9 (voice, Computer Use, robotics), followed by a merged summary and 12 merged exercises. - Old chapters 6/7/8 shift to 7/8/9; the old chapter 9 file is consumed by the new chapter 6. Chapter 4 keeps the three Agent-invoked tool categories and hands the event-triggered and user-communication tools to Chapter 6. Mechanical parts, applied per edition with language-specific label patterns - Cross-chapter references rotate 6→7→8→9→6 (including Arabic ordinal words and Traditional Chinese numerals). - Figures, experiments, tables and footnote IDs renumbered with one global map: fig4-2..4-6 → fig6-1..6-5, fig9-N → fig6-(N+5), fig4-7..4-9 → fig4-2..4-4, and 6→7, 7→8, 8→9 for the shifted chapters; 56 SVGs renamed per edition. - Companion code links (`../chapterN/`) rotate with the code directories. New prose written for every edition - Chapter title, opening, the "Two Axes: Modality and Timing" section with both tables, the async section heading and its lead-in, the chapter summary, and the twelfth exercise plus its reference answer. - Chapter 4's roadmap paragraph, tool-classification paragraph and summary are rewritten to hand off the two event-driven tool categories. - introduction and afterword: the part structure becomes 2–6 / 7–9 / 10, the chapter-6 bullet is rewritten and moved into place, the old multimodal bullet is retired, and the reading paths and prerequisite chapter numbers follow. - reference-answers: Chapter 4 drops to four answers, the three that moved plus the eight from the old chapter 9 and one new answer form the chapter 6 section, which is repositioned accordingly. Also fixes seven stale companion links in the Chinese edition that #856 left pointing at the pre-rotation directories. Verified: for all 12 editions × 10 chapters, the experiment inventory, figure inventory and order, and thought-question count match the Chinese exactly; no dangling footnotes, no missing image files, no unbalanced code fences. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
main squash-merged #858, whose content this branch already contains; the tree is kept as-is so only the Chapter 6 restructure remains as the diff. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…three-layer skeleton The book carried two rival ways of classifying defenses. Chapter 1 sorted guardrails by position in the request lifecycle (input side / execution side / output side); chapters 2 and 5 sorted them by where the trust boundary is enforced (context layer / execution layer / data layer). Chapter 2 already said "this confirms the layered-defense principle from Chapter 1" and used the context/execution vocabulary — a principle Chapter 1 did not actually state. Adopt the second taxonomy book-wide and front it in Chapter 1: - Chapter 1's "Types of Guardrails" is rewritten around **context layer / execution layer / data layer**, ordered not by lifecycle position but by how hard each layer is to bypass. The existing material is preserved and reassigned: the four input-side mechanisms become the context layer; tool risk rating stays at the execution layer; output checks join it, since a reply to the user is itself an action (Chapter 4 classifies it as a user-communication tool). Two points are added — the context layer's structural ceiling (an Agent inside the attacked context can hardly tell it has been injected, so this layer lowers success rates but gives no guarantee), and the data layer, which holds even when the layers above it fail. - Chapter 1's pointers now name the layer each chapter develops, and its summary states the skeleton instead of an unrelated five-level list. - Chapters 4, 5 and 9 are hung on the same tree: the MCP mitigations are split across the context and execution layers, the Coding Agent defense narrative and the "move the trust boundary down to the data layer" argument are named as layers of the Chapter 1 skeleton, and the self-evolution safety gate is tied to the same reasoning — a real guarantee must come from a layer the modifier cannot reach. The stray "execution side" wording in Chapter 4 is aligned to "execution layer". Applied to the Chinese edition and all 12 translations. Verified that the experiment, figure and thought-question inventories still match the Chinese across all 13 editions, with no dangling footnotes, missing images or unbalanced code fences. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
main squash-merged #859, whose content this branch already contains; the tree is kept as-is so only the security-taxonomy change remains as the diff. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Implements item 7 of the structural review: "安全有两套互不相认的分类法".
The problem
The book carried two rival ways of classifying defenses:
The giveaway: Chapter 2 already said "this confirms the layered-defense principle from Chapter 1" and used the context/execution vocabulary — a principle Chapter 1 never actually stated.
The fix
Adopt the second taxonomy book-wide and front it in Chapter 1.
Chapter 1's "Types of Guardrails" is rewritten around context layer / execution layer / data layer, ordered not by lifecycle position but by how hard each layer is to bypass. All existing material is preserved and reassigned:
Two substantive additions:
Hanging the rest of the book on the same tree:
Scope
Applied to the Chinese edition and all 12 translations in one PR, so no edition is left with the old taxonomy.
Verified: experiment, figure and thought-question inventories still match the Chinese across all 13 editions; no dangling footnotes, missing images, or unbalanced code fences.
🤖 Generated with Claude Code