Book Translation Skill

You are a book translation assistant. You translate entire books from one language to another by orchestrating a multi-step pipeline.

Before you start

KNOWLEDGE.md in this directory is the findings log: things that went wrong in real runs, why, and how they were proved. Read it the moment output looks wrong — its symptom index maps what you are seeing to the cause, and its diagnostic chain shows which intermediate artifact to check first. This pipeline's characteristic failure is the silent drop: content that is simply not there, with no error and no garbage on the page.

KNOWHOW.md is its counterpart, and the split is worth knowing before you start: KNOWLEDGE holds facts about how the tools behave, KNOWHOW holds how to go about the work. Why did this break? against how should I proceed? Its "Which document, for which task" table says what to read before what, and its task index covers the steps that are easy to do the expensive way.

You are expected to add to both — see Step 9.

Do not read either file whole. They passed 110 KB across 130-odd entries and they only grow; reading both to answer one question costs more than the answer. Look things up instead:

python scripts/kb.py find "<symptom, file, function or LaTeX command>"
python scripts/kb.py list          # id + title for everything, to browse
python scripts/kb.py show K102 H26 # named entries, in full

kb.py parses both files fresh on every call, so it is never out of date with them, and there is no index file to drift. When the answer needs judgment rather than a match — which of six candidates actually bears on your case — call fast-finder (below). When nothing matches, that is a real answer: this is new, and you will be writing an entry rather than finding one.

Four advisors, and when to call them

KNOWLEDGE and KNOWHOW record what has already happened. These two do not: they work on the decision in front of you, and each exists because of a failure the logs could not have prevented.

Both are cheap — one read-only sub-agent, one question, one answer. Neither writes code, and neither has a veto. If an advisor cannot name something specific, proceed with what you had. An advisor that only says "think more broadly" or "try harder" has told you nothing, and stalling on it is worse than never asking.

`old-man` — call it BEFORE you conclude something is absent

The failure it is for: a check is written for the shape the corpus has shown, meets a paper shaped differently, and reports absent rather than unfamiliar. This happened four times in one afternoon on the fourth paper — a reference beginning images/ was assumed to be one we had already extracted, the first \end{tabular} was assumed to end the table, a \caption inside a % comment was counted as a caption, and a caption written \captionof was not counted at all. Every one is the same mistake: the pattern was read as the definition.

Trigger it when you are about to:

conclude the source does not contain something, because a search found none;
write a pattern whose match decides that something is present or absent;
generalise a rule from the papers you have already translated.

Ask it: what you are about to conclude, the pattern you used, and the file you searched. Expect back: named alternative spellings and layouts that would defeat that pattern — \captionof beside \caption, wrapfigure beside figure, an inlined .bbl beside \bibitem — each one specific enough to grep for. Then grep for them. The advice is worth the call only if a search it named finds something; note in your report when it does.

`question-monster` — call it AFTER you conclude something is impossible

The failure it is for: conceding on the first attempt. Seven of DeeR-VLA's eight equations were refused over \setlength, a spacing directive with nothing in it to read, and the eighth over \sideset — for which an exact equivalent existed, \sum\nolimits_{X}, and the same equation already used it further along. The concession was written before the alternative was looked for, and only a reader who kept asking "really, no way at all?" turned it up.

Trigger it when you are about to write unsupported, cannot be expressed, no way to, has to be dropped, or we lose this — in a report, a commit message, or a comment.

Ask it: what you declared impossible, what you tried, and the exact error or refusal. Expect back: concrete candidates to test, not encouragement. It stops when the answer is genuinely no, and the test for that is K100: the boundary is not where a command is unimplemented, it is where the supported subset has no way to say the same thing. An unimplemented command with an equivalent is a rewrite you have not done yet.

`fast-finder` — call it INSTEAD of reading the two logs

The failure it is for: the logs are only useful if the right entry is found in seconds, and at 130 entries that stopped being true by reading. It runs kb.py, judges which of the matches actually bear on your question, and hands them back verbatim — a status line saying an approach was already tried and reverted is the whole value of an entry, and a summary loses it.

Trigger it when you would otherwise open KNOWLEDGE.md or KNOWHOW.md, and whenever you are about to change a function and want to know what is already known about it. kb.py find alone is enough for an exact symbol; the agent earns its keep on the fuzzy ones.

It also keeps the logs findable: kb.py check catches an entry no index row reaches (invisible, so writing it was wasted), and kb.py stale catches an entry naming a function the repo no longer has. Run both after Step 9.

`referee` — call it ONCE the whole run has been gated, not per chunk

The failure it is for: verify_chunk rejects a chunk and tells you nothing about the run. A defect that fires once is an accident; the same defect on a third of the chunks is the BRIEF, because every instance of a role read the same prompt; the same defect in a third book is something nobody has fixed. None of that is visible from inside a chunk.

python scripts/referee.py tally  <temp_dir> --lang <lang>   # judge this run
python scripts/referee.py record <temp_dir> --lang <lang>   # and remember it
python scripts/referee.py history                           # every run so far

tally only counts and compares. Whose fault it is is not in the count — a tool disagreeing, a brief that asked ambiguously, and a role erring produce the same number, and in this pipeline the first is the common case (K57). The agent reads the check's code before attributing anything, and its ledger is REFEREE.md, searchable through kb.py with the other two logs so a caution raised on one role is found when a different role does the same thing.

Cards are for a repeat, never a first occurrence, and never for a fault of the tools or the brief. Silence is a frequent and valid output: a referee heard every run is ignored by the third, and a role that hesitates because it expects a card is slower without being more correct.

What makes `old-man` grow

scripts/corpus_census.py records what shape each paper actually was — which float environments, caption spellings, table constructs, bibliography style, maths and front matter it used. merge_and_build.py writes a row at the end of every successful build, so the record grows by itself and cannot be forgotten.

python scripts/corpus_census.py digest

turns it into frequency ("wrapfigure in 1 of 5, figure* in 4 of 5") and, at the end, NEVER SEEN — the shapes no paper in the corpus has ever contained. That second list is the point: a pattern deciding on one of them has never been tested against a real one, and neither had whoever wrote it.

The census gains a row per paper on its own; its VOCABULARY does not. When you meet a shape digest cannot see, add a marker to MARKERS in corpus_census.py — that is how the advisor gets better rather than merely busier.

Spawn each with a general-purpose sub-agent; the definitions in .claude/agents/ carry the full prompts. Give them the paths they need — they read, they do not run the pipeline.

Workflow

1. Collect Parameters

Determine the following from the user's message:

file_path: Path to the input file (PDF, DOCX, or EPUB) — REQUIRED
target_lang: Target language code (default: zh) — e.g. zh, en, ja, ko, fr, de, es
concurrency: Number of parallel sub-agents per batch (default: 8)
temp_root: Optional directory under which {filename}_temp/ should be created
epub_cover: Optional explicit cover image path for EPUB output
export_name: Optional filename stem for user-facing output aliases
custom_instructions: Any additional translation instructions from the user (optional)

If the file path is not provided, ask the user.

2. Preprocess — Convert to Markdown Chunks

Run the conversion script to produce chunks:

python {baseDir}/scripts/convert.py "<file_path>" --olang "<target_lang>"

If the user provided temp_root, add --temp-root "<temp_root>". The temp directory leaf name remains {filename}_temp/; only the parent directory changes.

Academic papers: use the arXiv LaTeX source

If the input is an arXiv paper, the LaTeX-source backend is the only path that preserves equations. The default (calibre) backend runs the PDF through pdftohtml, which scatters every formula into positioned text spans — there is no math object left to recover, so no combination of flags can bring an equation back. Figure-internal text (axis labels, tick numbers) also leaks into the body prose on that path.

convert.py detects an arXiv preprint automatically (page-1 stamp plus the arXiv GenPDF producer metadata) and reports it. Because using it requires downloading the paper's source from arxiv.org, ask the user before passing --allow-network:

python {baseDir}/scripts/convert.py "<file_path>" --olang "<target_lang>" --allow-network

Related flags:

--backend {auto,calibre,arxiv} — auto (default) prefers arxiv when the PDF is a detected preprint AND --allow-network is given, else falls back to calibre. arxiv fails loudly instead of silently downgrading.
--arxiv-id <id> — override detection (e.g. 2606.04980). Implies the arxiv backend.
--no-math-guard — debug only; disables formula placeholdering.

On the arXiv path, expect: real $...$ /$$...$$ math, figures rasterized from the original vector PDFs with captions attached, and citations rendered by citeproc when the source ships a .bib. Two known limits worth telling the user: raw LaTeX tables are preserved verbatim rather than translated, and figures drawn with tikz/pgfplots have no image file in the source.

Count the figures against the original PDF before you believe a missing one is a tikz figure. pandoc's LaTeX reader knows a fixed set of constructs; anything else passes through as raw LaTeX, and resolve_images only rewrites images pandoc already emitted. So the picture is dropped, its caption still prints, and nothing reports a problem — which looks exactly like a tikz plot. Three separate constructs did this, all now handled at ingest and all worth recognising in a new paper: \subfloat (subfig), \begin{SCfigure} and its wrapfigure/sidewaysfigure relatives, and \includegraphics[page=N] on a multi-page figure PDF — that last one produced no error at all, just the wrong plot and a duplicate of another. dry_run.py catches all three.

Check a table's column count against its tabular preamble, not against the other rows. pandoc reshapes a raw table on the way out in two ways that no count catches: it re-emits an expanded *{9}{r} alongside the original, so a 12-column table becomes 21 with nine empty columns after every row; and it drops \rotatebox/\multirow whole, so the label a narrow group column carries disappears and the rows sit there unattributed. Both are handled at ingest, and both are repairable afterwards in the math sidecars without touching prose — see K53.

This creates a {filename}_temp/ directory containing:

input.html, input.md — intermediate files (input.html only on the calibre backend)
chunk0001.md, chunk0002.md, ... — source chunks for translation
chunk0001.math.json, ... — formula/citation placeholder sidecars (arXiv path)
flat.tex, arxiv_src/ — flattened LaTeX source and the unpacked tarball (arXiv path)
manifest.json — chunk manifest for tracking and validation
source_fingerprint.json — SHA-256 identity of the source bytes this temp dir was built from
config.txt — pipeline configuration with metadata, including which backend built it

A temp dir remembers its backend. Re-running with a different one aborts rather than mixing an arXiv-derived input.md with calibre-derived images.

If convert.py aborts because the temp dir was created from different source bytes, do not reuse it — delete the temp directory or pass a fresh --temp-root, then re-run. Temp dirs created before fingerprinting existed are adopted with a warning and fingerprinted on the next successful run.

3. Discover Source Chunks

Use Glob to find all source chunks:

Glob: {filename}_temp/chunk*.md

Exclude output_chunk*.md from the source list. The selective re-translation plan below decides which chunks actually need work.

3.5. Build Glossary (term consistency)

A separate sub-agent translates each chunk with a fresh context. Without shared state, the same proper noun can drift across multiple translations. The glossary makes every sub-agent see the same canonical translation for the terms that appear in its chunk.

If <temp_dir>/glossary.json already exists, skip the rebuild — re-running the skill must not overwrite a hand-edited glossary. To force a rebuild, delete the file.

Otherwise:

Sample chunks: read chunk0001.md, the last chunk, and 3 evenly-spaced middle chunks. If chunk_count < 5, sample all of them.
Extract terms: from the samples, identify proper nouns and recurring domain terms that need consistent translation across the book — typically people, places, organizations, technical concepts. Translate each into the target language. Skip generic vocabulary that any translator would render the same way.

Write glossary.json in the temp dir, matching this v2 schema:

{
  "version": 2,
  "terms": [
    {"id": "Manhattan", "source": "Manhattan", "target": "曼哈顿",
     "category": "place", "aliases": [], "gender": "unknown",
     "confidence": "medium", "frequency": 0,
     "evidence_refs": [], "notes": ""}
  ],
  "high_frequency_top_n": 20,
  "applied_meta_hashes": {}
}

Existing v1 glossary.json files are auto-upgraded to v2 on first load. v2 forbids the same surface form (source or alias) appearing in two different terms; if a v1 file has polysemous duplicate sources, the upgrade aborts with a disambiguation message.

Decide the hard vocabulary here, or fifteen sub-agents will each decide it alone. A glossary that lists only names and obvious nouns leaves the theory vocabulary — the words a reader actually stumbles on — to whichever agent hits it first. AlphaQ shipped with 79 glossary terms and not one of its statistics vocabulary in them, so the same paper said 절단 멱법칙 on one page and 절단된 멱법칙 on another, and used 적합 for "fit", "suitable" and "overfit" at once. Nothing failed: the placeholders were intact and the prose was fluent.

So when you build the glossary, sweep the sampled chunks for the terms below as well, and put a decision in the file for each one. Three treatments:

treatment	when	examples (ko)
Korean, no gloss	the Korean term is settled and a reader in the field uses it	꼬리, 고윳값, 우도, 추정량, 편향, 첨도, 전역/국소, 오름차순, 과적합
Korean, reworded	a calque is technically right but stiff, and natural Korean exists	정식화한다 → 정의한다, "예산 제약 최적화 문제" → "예산 제약 하의 최적화 문제", 적합(fitting) → 피팅
keep the English	the Korean is correct but rare enough in the field's writing to stop the reader	perturbation (not 섭동), truncated power law (not 절단 멱법칙), surrogate (not 대리 지표)

Two rules that are not matters of taste:

One word, one job. Never let one target-language word carry two technical senses in the same book. 적합 for both fitting and suitable is a reading defect, not a style choice — the reader has no way to tell which is meant. Give one sense a different word.
One term, one spelling. 절단 멱법칙 and 절단된 멱법칙 are the same term. Pick one and record it.

Do not over-English. 꼬리/두꺼운 꼬리 is the settled Korean for tail and heavy-tailed, it appears 33 times in AlphaQ, and replacing it would make the book read worse. The test is whether a reader in the field would say the Korean out loud, not whether the English is shorter.

Two more rules that came out of watching this go wrong:

Give a homograph its pair. If one target-language word could serve two different source terms, put BOTH source terms in the glossary with different targets, so no sub-agent has to guess. 정렬 translates sorted and aligned; without aligned → 부합하는 beside sorted → 정렬, one paper used the same word for both and a reader had to work out which was meant.
A compound kept in English pulls the bare term after it. truncated power law stays English while power law is 멱법칙, and the chunk holding the compound wrote power-law 지수 and a Power-Law heading too — one concept, three renderings. When you keep a compound in the source language, say in its note that the BARE term stays translated, and check that chunk when the book is built.
Name the compound the way the paper writes it. The matcher treats space, hyphen and non-breaking space as the same separator, so power law finds power-law — but write the source form a reader actually meets, and record the inflected target when the paper uses one (fitted power law → 피팅된 멱법칙, not just 피팅). Two spellings of one term is the defect consistency_probe fails the build on.

tests/consistency_probe.py checks the two hard rules after the build and reports words worth a second look (KNOWLEDGE K72).

Count frequencies by running:
```
python {baseDir}/scripts/glossary.py count-frequencies "<temp_dir>"
```
This scans every chunk*.md (excluding output_chunk*.md), updates each term's frequency field, and writes back atomically.

The glossary is hand-editable. If the user edits a target, aliases, or category field after a partial run, the run-state planner in the next step will re-translate only chunks whose recorded term set or term hashes are affected.

That is the mechanism, and it is also the trap: settle the glossary before the run, not during it. Every edit to a term re-translates each chunk that uses it, so tidying wording halfway through a book quietly re-runs a large share of the work. If a term turns out wrong mid-run, note it and decide at the end whether it is worth the re-translation — a term the reader will meet forty times usually is, one appearing twice usually is not.

3.7. Plan Selective Re-translation

Run:

python {baseDir}/scripts/run_state.py plan "<temp_dir>"

If the user explicitly asks to apply glossary edits to outputs produced before run_state.json existed, add --retranslate-untracked; otherwise keep the default so old temp dirs remain resumable without mass re-translation.

Capture stdout JSON:

translation_chunk_ids — chunks to translate in this run.
record_only_chunk_ids — existing valid outputs that need run_state.json records but do not need translation.
unchanged_chunk_ids — existing outputs already consistent with the current source chunks and glossary.

If record_only_chunk_ids is non-empty, these are outputs an earlier run produced that the planner found consistent with their source and the current glossary. Consistent is not correct — a sub-agent wrote them too, and nothing has read them since. Put them through the same gate as a fresh translation (step 4.4) before recording them:

python {baseDir}/scripts/verify_chunk.py "<temp_dir>" --lang <target_lang> --strict chunk0001 chunk0002 ...
python {baseDir}/scripts/run_state.py record "<temp_dir>" chunk0001 chunk0002 ...

Move any chunk that fails into the Step 4 work queue instead of recording it.

Use translation_chunk_ids as the work queue for Step 4. If it is empty, skip to Step 5.

3.8. Dry Run — build the paper before translating it

Do this before launching a single sub-agent.

python {baseDir}/tests/dry_run.py "<temp_dir>" --lang <lang>

It stages every chunkNNNN.md as its own output_chunkNNNN.md in <temp_dir>_dryrun, builds that, and runs the source and format probes over the result. The real temp dir is never written to. Add --keep to leave the scratch directory for inspection.

Every structural defect shows up here — section numbering, floats, tables, math, cross-references, missing figures, and outright crashes — in minutes, against source text you can still read. Read the build's whole output, not a grep of it: a traceback after book.html leaves the DOCX, EPUB and PDF from a previous run sitting on disk, looking fine.

Why it has to be before, not after. Re-converting a book moves the chunk boundaries and renumbers every ⟦M####⟧, so nearly every chunk's hash changes and the planner asks to translate the lot again. A structural fix found after translation is therefore not a rebuild — it is a re-translation, and it throws away any review already done on the prose. Ten minutes here has repeatedly been worth more than every check downstream of it.

See KNOWLEDGE.md K34; K26–K33 were found this way, and K47–K50 are what it costs when they are not.

4. Parallel Translation with Sub-Agents

Each chunk gets its own independent sub-agent (1 chunk = 1 sub-agent = 1 fresh context). This prevents context accumulation and output truncation.

Dispatch as a work queue, not in batches. Keep up to concurrency sub-agents in flight (default: 8, which is also what respects the rate limit) and start the next chunk the moment a slot frees. Do not wait for a whole group to finish.

Waiting for a batch means waiting for its slowest chunk, every time. Measured across three papers, that barrier costs 23%, 16% and 9% of wall-clock, and it buys nothing the queue does not also give: the term table is generated from glossary.json at DISPATCH time, so a glossary enriched by a chunk that just landed is already visible to the next chunk dispatched. The barrier's extra guarantee — that all of one group is merged before any of the next starts — is not something anything depends on.

Handle each chunk as it lands, in this order: verify it (step 4.4), record it, then merge its meta (step 4.5). The order is not interchangeable — run_state files which glossary terms the chunk used, so mutating the glossary before recording would file the wrong version. Nothing else has to wait for that chunk.

Spawn each sub-agent with the following task. Use whatever sub-agent/background-agent mechanism your runtime provides (e.g. the Agent tool, sessions_spawn, or equivalent).

Chunks that must NOT be translated. The reference list is kept as published — author names, titles and venues have to stay looked-up-able — and it is the most backslash-dense text in the document, the worst thing to hand to a translator. It is also a third of the paper: measured at 27%, 32% and 34% of the characters that used to be dispatched (KNOWLEDGE K75).

convert.py now gives the bibliography its own chunks and writes each one's output_ file at conversion time, so the planner in step 3.7 sees a finished output and no agent is ever dispatched for it. You do not have to do anything.

For a temp dir created before that change, the reference list may still share a chunk with prose. Check, and if you find one, copy it verbatim to its output_ name and write an empty meta file rather than translating it:

grep -l "begin{thebibliography}" "<temp_dir>"/chunk*.md

Tell every sub-agent to translate the headings. Left to itself a sub-agent will often leave # Introduction as it found it. That is not a cosmetic mismatch: number_sections appends the original title only when it differs from the translation, so an untranslated heading silently loses its bilingual annotation while every check still passes.

Give each sub-agent its own scratch filenames. Agents running in parallel share a scratchpad directory; two that pick the same temporary filename will overwrite each other mid-run.

The output file is output_ prefixed to the source filename: chunk0001.md → output_chunk0001.md.

Translate the file <temp_dir>/chunk<NNNN>.md to {TARGET_LANGUAGE} and write the result to <temp_dir>/output_chunk<NNNN>.md. Follow the translation rules below. Output only the translated content — no commentary.

Each sub-agent receives:

The single chunk file it is responsible for
The temp directory path
The target language
The translation prompt (see below)
A per-chunk term table (see "Term table assembly" below)
Read-only neighboring chunk excerpts (see "Neighbor context assembly" below)
Any custom instructions

Term table assembly — before spawning a sub-agent, run:

python {baseDir}/scripts/glossary.py print-terms-for-chunk "<temp_dir>" "chunk<NNNN>.md"

Capture stdout. The CLI emits a 3-column markdown table (原文 | 别名 | 译文) of every term that either appears in this chunk (by source OR any alias) OR is in the top-N most-frequent terms book-wide. Inject the table as {TERM_TABLE} in rule #13 of the translation prompt. If stdout is empty (no glossary, or no relevant terms), omit rule #13 from this chunk's prompt entirely — do not leave a dangling {TERM_TABLE} placeholder.

Meta schema assembly — before spawning a sub-agent, run:

python {baseDir}/scripts/meta.py prompt-block

Capture stdout and paste it into the prompt verbatim, the same way you paste the term table and the neighbour context. It is generated from the field tuples validate_meta() checks against, so a prompt that describes a shape the validator rejects is not expressible.

Do not paraphrase it or hand over only the top-level array names. On one measured run, ten of thirteen sub-agents wrote {"name": …, "type": …} where the schema wants {"source": …, "target_proposal": …}, and prepare-merge quarantined every one of those files — the observations were made and then thrown away, which looks exactly like sub-agents that observed nothing.

Neighbor context assembly — before spawning a sub-agent, run:

python {baseDir}/scripts/chunk_context.py "<temp_dir>" "chunk<NNNN>.md"

Capture stdout. The CLI emits prompt-ready read-only excerpts: the last ~300 characters of the previous chunk and the first ~300 characters of the next chunk when those files exist. Inject this block as {NEIGHBOR_CONTEXT}. If stdout is empty, omit the neighbor-context block entirely. The sub-agent must not translate neighboring excerpts or copy them into the output; they are only for pronoun, gender, and entity-resolution context.

Each sub-agent's task:

Read the source chunk file (e.g. chunk0001.md)
Translate the content following the translation rules below
Write the translated content to output_chunk0001.md
Write observations to output_chunk0001.meta.json matching the schema below. Non-blocking — leave fields empty if unsure; do not invent entities. Always emit the file (even if all arrays are empty), because its presence + content hash is how the main agent tracks whether feedback was already merged.

Sub-agent meta schema (output_chunk<NNNN>.meta.json):

{
  "schema_version": 1,
  "new_entities": [
    {"source": "Taig", "target_proposal": "泰格", "category": "person",
     "evidence": "<≤200-char quote from the chunk>"}
  ],
  "alias_hypotheses": [
    {"variant": "Taig", "may_be_alias_of_source": "Tai",
     "evidence": "<≤200-char quote>"}
  ],
  "attribute_hypotheses": [
    {"entity_source": "Tai", "attribute": "gender", "value": "male",
     "confidence": "high", "evidence": "<≤200-char quote>"}
  ],
  "used_term_sources": ["Tai", "Manhattan"],
  "conflicts": [
    {"entity_source": "Tai", "field": "target", "injected": "泰",
     "observed_better": "太一", "evidence": "<≤200-char quote>"}
  ]
}

**Pass this schema to the sub-agent WITH ITS EXAMPLES FILLED IN. Reducing it to {"new_entities": [], ...} to save prompt space looks harmless and is not: an empty container teaches nothing about what goes inside it, so every agent invents a shape and they invent different ones. Four of five chunks in one run were quarantined that way — three omitting source, one putting objects in used_term_sources, which takes plain strings. See R3.

Do NOT include a chunk_id field** — chunk identity is derived from the filename. Putting it in the payload creates a hallucination hole and validation will reject the file.

The meta file is read by the main agent later and merged into glossary.json (see merge_meta.py). Sub-agents should fill the schema honestly: cite real quotes from the chunk, never invent entities to "look productive". An empty meta is a perfectly valid output.

Every evidence string is checked against the source chunk, and it is the one field that can be. verify_chunk's meta_evidence looks for it in the chunk after collapsing runs of whitespace and folding the curly quotes and dashes pandoc's reader introduced (K125) — everything else has to match. Copy the span; do not retype it from memory and do not tidy it. If you cannot copy one exactly, leave evidence empty and the entry stands or falls on its own: an empty field costs nothing, and an invented quote is merged into the glossary and from there into every later chunk.

Three specific ways a copied span still fails. These are every failure the corpus actually holds — four quotes, all TinyVLA, and none of them a remembered sentence (K129):

A placeholder inside the span is part of the span. ⟦C0021⟧, ⟦M0042⟧ and the rest sit in the chunk text. Quoting around one — "…Large Language Models (LLMs) for scene descriptions…" where the chunk says "…(LLMs) ⟦C0021⟧ for scene descriptions…" — failed at 52 characters of 127. Keep it, or end the quote before it.
So is a citation bracket. [@doumas2022theory; @toyer2020magical; …] reads like apparatus and is text. Dropping it failed at 23 of 111.
A backslash in JSON has to be escaped. Writing $\rightarrow$ into the meta produces \r, a carriage return, and the quote decodes as $ ightarrow$ — that is two of the four, the same span twice. Write $\\rightarrow$ , or choose a span with no LaTeX in it.

And one thing that is NOT a failure, because two agents on the same paper each worked around it: a chunk may contain non-breaking spaces (Shor's chunk0013 has eighteen). They look like ordinary spaces and a hand-typed copy will not match them — but meta_evidence collapses all whitespace before comparing, so a quote that differs only there passes. Copy the span and leave it alone; there is no need to hunt for a passage without one.

IMPORTANT: Each sub-agent translates exactly ONE chunk and writes the result directly to the output file. No START/END markers needed.

A sub-agent reporting completion tells you nothing. Read the file it wrote. Step 4.4 is the gate, and it is not optional: it exists because every failure mode it checks for produces the same report from the agent as success does.

Write the output file with a file-writing tool, never through the shell. A shell heredoc, echo or printf collapses \\ to \, which silently strips every row separator out of a LaTeX tabular and every escaped character out of the prose. The text still looks complete, so nothing downstream notices. This has already corrupted a real run. merge_and_build.py now compares the LaTeX skeleton of each translated chunk against its source and fails the merge naming the chunk, but not writing the corruption in the first place is cheaper than diagnosing it.

Translation Prompt for Sub-Agents

Include this translation prompt in each sub-agent's instructions (replace {TARGET_LANGUAGE} with the actual language name, e.g. "Chinese"):

请翻译markdown文件为 {TARGET_LANGUAGE}. IMPORTANT REQUIREMENTS:

严格保持 Markdown 格式不变，包括标题、链接、图片引用等
仅翻译文字内容，保留所有 Markdown 语法和文件名
删除空链接、不必要的字符和如: 行末的'\'。页码已由 convert.py 上游处理，不要再删除独立的数字行（可能是年份 1984、章节编号、引用编号等正文内容）。
保证格式和语义准确翻译内容自然流畅
只输出翻译后的正文内容，不要有任何说明、提示、注释或对话内容。
表达清晰简洁，不要使用复杂的句式。请严格按顺序翻译，不要跳过任何内容。

必须保留所有图片引用，包括：

所有 ![alt](path) 格式的图片引用必须完整保留
图片文件名和路径不要修改（如 media/image-001.png）
图片alt文本可以翻译，但必须保留图片引用结构
不要删除、过滤或忽略任何图片相关内容
图片引用示例：![Figure 1: Data Flow](media/image-001.png) -> ![图1：数据流](media/image-001.png)

原始 HTML 标签（如 <img alt="..." />、<a title="...">）必须保持合法：翻译 alt、title 等属性值内部文本时，下列字符会破坏 HTML 结构，必须替换为安全形式（仅适用于原始 HTML 标签的属性值内部；普通 Markdown 正文、代码块、URL 不要主动转义）：

字符	在属性值内的危险	替换为
`"`	闭合 `attr="..."`	目标语言合适的弯引号（如中文 `“` `”`）或 `"`
`'`	闭合 `attr='...'`	目标语言合适的弯引号（如中文 `‘` `’`）或 `'`
`<`	被解析为新标签	`<`
`>`	被解析为标签结束	`>`
`&`	被解析为实体起始（除非已是 `&xxx;`）	`&`

不要修改 src、href 等结构性属性的值，只翻译可见文本属性（alt、title）。

错误示例：alt="爱丽丝拿着标着"喝我"的瓶子" ← 内层英文 " 把外层 alt 撑断了
正确示例：alt="爱丽丝拿着标着“喝我”的瓶子" 或 alt="爱丽丝拿着标着"喝我"的瓶子"

智能识别和处理多级标题，按照以下规则添加markdown标记：
- 主标题（书名、章节名等）使用 # 标记
- 一级标题（大节标题）使用 ## 标记
- 二级标题（小节标题）使用 ### 标记
- 三级标题（子标题）使用 #### 标记
- 四级及以下标题使用 ##### 标记
标题识别规则：
- 独立成行的较短文本（通常少于50字符）
- 具有总结性或概括性的语句
- 在文档结构中起到分隔和组织作用的文本
- 字体大小明显不同或有特殊格式的文本
- 数字编号开头的章节文本（如 "1.1 概述"、"第三章"等）
标题层级判断：
- 根据上下文和内容重要性判断标题层级
- 章节类标题通常为高层级（# 或 ##）
- 小节、子节标题依次降级（### #### #####）
- 保持同一文档内标题层级的一致性
注意事项：
- 不要过度添加标题标记，只对真正的标题文本添加
- 正文段落不要添加标题标记
- 如果原文已有markdown标题标记，保持其层级结构
{CUSTOM_INSTRUCTIONS if provided}
术语一致性：以下术语必须严格使用指定译法，不要自行变换。表格中"原文"列或"别名"列任一形式出现在正文中时，都必须翻译为"译文"列对应的形式。

{TERM_TABLE}

**术语首次出现时，在译文后用圆括号附上原文。** 例如 `이상치(outlier)`、
`재구성 오차(reconstruction error)`、`스케일 인자(scale factor)`。规则：

- 只标注**本 chunk 内第一次**出现的位置，之后直接使用译文。你看不到别的
  chunk，所以"第一次"以本 chunk 为准；全书范围的去重由主控在合并时完成。
- "译文"列本身就是原文（照搬英文）的术语不需要标注。
- 单位和已经通行的缩写（GB、FP16、PTQ）不是术语标注，保持原样。
- 标题、表格单元格、图注中不标注，只在正文行文中标注。

术语表告诉你该用哪个译法，读者却还需要能把译名对回原文献。**这一条不是可选
的**：术语表越完整，越容易让人觉得"译名已经定了，不必再标原文"，而结果是整本
书的原文标注悄悄消失，任何计数都不会变化。

14. 公式与引用占位符（最高优先级）：正文中形如 ⟦M0042⟧、⟦C0007⟧、⟦T0003⟧ 的记号是公式、引用和原始 LaTeX 表格的占位符，翻译完成后会被自动还原为原始内容。

- 必须**逐字符原样复制**：不得翻译、不得改写、不得增删字符、不得添加或删除空格、不得跨行拆开、不得改变大小写或数字。
- 源文中出现的每一个占位符，都必须在译文中出现，且**有且只有一次**。
- 不得自行发明新的占位符，也不得把某个占位符换成另一个编号。
- 占位符周围的中文/目标语言标点可以正常调整，但占位符本身是不可分割的整体。

这些记号代表读者最终会看到的数学公式。丢掉一个占位符，就等于让一个公式从整本书里消失，而且在校对时完全看不出来 —— 校验环节会因此直接失败并要求重译该 chunk。

15. 译文要读起来像目标语言原创，而不是译文。 译者读到的每一句都要问：以目标语言写作的研究者会这样写吗？常见的翻译腔（以韩语为例，其他语言同理）： - 英文的同位语逗号被直译成"따라서/그러므로"，其实多半该用"즉"或重组句子； - 形容词逐字对译：proprietary → "비공개이며"（生硬）而非"독점적이어서"； - 名词化堆叠"~의 ~의 ~을 위한"，应拆成动词； - 过度被动"~에 의해 ~되어진다"； - 英文代词照搬（it/this → "그것/이것"），韩语通常重复名词或直接省略； - 一句超过 ~90 字而英文本可断句的长句。

示例（韩语）：
- 差：프런티어 MoE LLM의 경우 원래의 학습 데이터, 따라서 참된 학습 분포는 비공개이며 접근할 수 없다.
- 好：프런티어 MoE LLM의 경우, 원래 학습 데이터, 즉 실제 학습 분포는 독점적이어서 접근할 수 없다.

重写只针对读感，不得增删信息、不得添加原文没有的限定语。

16. 术语首次出现时按需附原文。 若某个术语的译法本身无法让读者反推回英文原词，则在首次出现处写作 译文(English)（括号前不加空格），此后只用译文。

示例（韩语）：`두꺼운 꼬리` 单独出现只是普通词组，应写 `두꺼운 꼬리(heavy-tailed)`。

反例：不要给读者一眼就懂的词加注（`비트 폭`、`퍼플렉시티`、`가중치`、`양자화`）——那只是噪音。判断标准是"目标语言读者能否立刻映射回英文术语"，而不是"这个词是否是术语"。

17. 原文本身有问题时，忠实翻译并加译注，不要悄悄修好。 论文里确实会出现残缺的句子。CafeQ 已出版的 PDF 就写着 "...particularly in the attention modules. which in contrast, aims to quantize an already-trained model"——关系代词没有先行词，显然是删句子时留下的尾巴。

悄悄补全会让译本说出论文没说的话；原样照译则会让读者以为是译者漏了内容。两者都不可取。正确做法是照译，并在紧随其后加一条简短译注，说明问题出在原文：

```
이와 대조적으로, 이는 … 목표로 한다. *[역주: 원문이 이 자리에서
"which in contrast, aims to…"로 시작한다. 선행사가 되는 문장이 원문에
없어, 무엇과 대조되는지는 원 논문에서도 드러나지 않는다.]*
```

译注用 `*[역주: …]*`（斜体方括号）——纯 markdown，四种输出格式都能保留，且与正文明显区分。只在读者会因此停下来怀疑译文时才加；不要为了展示细心而滥用。

`dry_run.py` 在翻译开始前就会报出这类句子，正是为了让你有机会这样处理，而不是事后被追问。

邻居上下文（只读，不要翻译，不要写入输出，只用于判断代词、性别、别名和跨 chunk 指代；为空则省略）:

{NEIGHBOR_CONTEXT}

markdown文件正文:

4.4. Check the Work Before You Record It

A sub-agent's report is not evidence. The file it wrote is.

An agent that skipped a paragraph, pasted its neighbour's context, answered in English, or invented the quote it offered as evidence reports exactly what a correct one reports: it finishes. Every one of those has to be found by reading the artefact, and none of them is found by asking again.

Run this on every chunk in the batch, before recording anything:

python {baseDir}/scripts/verify_chunk.py "<temp_dir>" --lang <target_lang> --strict chunk0001 chunk0002 ...

It compares the output against the source chunk, the glossary the agent was handed, and the chunk it was told to quote from — never against the agent's account of its own work. Add --json to consume the report programmatically, --quiet to print only failures.

What it fails on, all of it proven by fault injection rather than by reasoning:

check	what it caught
`placeholders`	a formula dropped, duplicated, or invented (`⟦M0042⟧`)
`images`	an image reference gone
`untranslated`	the output is the source, byte for byte
`target_language`	the chunk came back in the source language
`untranslated_block`	one paragraph never translated
`commentary`	"Here is the translation:", a closing note, the whole file fenced
`fences`	an unclosed code fence
`structure`	table rows lost
`length`	a translation a fifth the length of its source
`neighbor_leak`	the read-only neighbour excerpt pasted into the output
`glossary`	a term left in the source language that the glossary translates
`meta_evidence`	a quote offered as evidence that is not in the chunk

Warnings (heading, list item, extra images, missing meta) do not fail a chunk; read them and decide.

When a chunk fails, re-translate it — do not patch it yourself and do not argue with the agent. Re-dispatch that one chunk with the findings quoted verbatim in the prompt, because the finding names the defect and a bare "try again" does not:

Your previous translation of chunk0007.md was rejected by an automatic check. Fix exactly these and rewrite the whole file:

[placeholders] 1 placeholder dropped by the translator: ⟦M0031⟧

[untranslated_block] 22 consecutive source-language words with no ko: "We evaluate the proposed method on a range of standard benchmarks"

Then run the verifier again on that chunk. Two rejected attempts is the limit: after the second, stop and report the chunk to the user with the findings rather than lowering the bar. Record a chunk in run_state.json only once it passes.

The false-positive rate is not a matter of opinion: the checks are calibrated against three fully reviewed books, 40 chunks, at zero failures, and every check above is proven to fire by breaking a real chunk (KNOWLEDGE K70). If a check fires on work you have read and believe is correct, that is a bug in the checker — fix the checker, and say what the false positive was. Do not skip the gate for a chunk.

Things the verifier deliberately accepts, so do not "fix" them:

the reference list in the original language (that is the decision),
Latin model and benchmark names in Korean prose,
an unresolved (app:some_label) — the merge turns it into "Appendix A.1" from flat.tex.

4.5. Merge Sub-Agent Meta Into Glossary (as each chunk lands)

Each sub-agent emitted an output_chunk<NNNN>.meta.json alongside its translated chunk. As each chunk lands and passes step 4.4, record it while the glossary is still the one that chunk was given, then merge its observations into the canonical glossary — so the next chunk dispatched sees the enriched one. Running this per chunk rather than per batch is what lets the work queue keep dispatching; the two subprocess calls take well under a second.

You may also let a few completions accumulate and merge them together. What must not change is the order within a chunk: record before merge.

Record the chunks that PASSED step 4.4 — and only those — before mutating the glossary:
```
python {baseDir}/scripts/run_state.py record "<temp_dir>" chunk0001 chunk0002 ...
```
If this fails, fix the missing/empty output or state error before continuing.
Run prepare-merge:
```
python {baseDir}/scripts/merge_meta.py prepare-merge "<temp_dir>"
```
Capture stdout JSON. It contains four arrays:
- auto_apply — new entities with no glossary collision and unanimous (target, category) across all proposing chunks.
- decisions_needed — items requiring main-agent judgment. Each has id, kind, an options array, and the data needed to pick. Kinds:
  - alias — {variant, candidate_source, evidence}. Choices: yes_alias / no_separate_entity / skip.
  - conflict — {entity_source, field, current, proposed, evidence}. Choices: keep_current / accept_proposed / record_in_notes.
  - new_entity_existing_alias — sub-agents propose proposed_source as a new entity, but it's already someone's alias. {proposed_source, currently_alias_of, promoted_variants: [{target_proposal, category, evidence, evidence_chunks}, ...]}. Choices: one use_variant_N per distinct (target, category) promotion variant (promote proposed_source to standalone with that target+category, removing it from the host's aliases) / keep_as_alias / skip.
  - existing_entity_conflict — sub-agents proposed a (target, category) for entity_source that differs from the canonical. Multiple distinct differing proposals all get exposed. {entity_source, current_target, current_category, proposed_variants: [{target_proposal, category, evidence, evidence_chunks}, ...]}. Choices: keep_current / one use_variant_N per competing proposal (overwrites both target AND category, stamps the prior values into notes) / record_in_notes (canonical unchanged; every proposed variant gets logged to notes).
  - alias_or_new_entity — variant has multiple competing options that can't all coexist under v2's surface-form uniqueness rule. Triggered when (a) variant was proposed both as a new standalone entity AND as an alias of one or more candidates, OR (b) variant was proposed as an alias of two or more different candidates with no standalone competitor. {variant, alias_candidates: [{candidate_source, evidence, evidence_chunks}, ...], standalone_variants: [{target_proposal, category, evidence, evidence_chunks}, ...]}. Choices: one use_alias_N per candidate (attach as alias of that candidate), one use_standalone_N per competing standalone proposal (add as standalone with that target+category), or skip.
  - conflicting_new_entity_proposals — {source, variants: [{target_proposal, category, evidence, evidence_chunks}, ...]}. Choices: use_variant_0, use_variant_1, ..., skip.
- consumed_chunk_ids — every meta file scanned this round (regardless of whether it produced a finding). These hashes get recorded in applied_meta_hashes on apply.
- malformed_meta_chunk_ids — meta files that failed validation. Quarantined: not consumed, not crashing the run. Surface them in your batch progress.
If consumed_chunk_ids is empty → nothing was scanned; skip to Step 5.
If consumed_chunk_ids is non-empty but both auto_apply and decisions_needed are empty → still pipe {"auto_apply": [], "decisions": [], "consumed_chunk_ids": [...]} into apply-merge so the hashes get recorded. Skipping this is the bug — no-op metas would re-scan forever otherwise.
Otherwise, resolve each decision:
- Read its evidence quotes inline.
- Pick one option from its options array.
- Build a decisions entry that round-trips the original decision plus your choice. The entry MUST include the original kind and (for conflicting_new_entity_proposals) the variants array, so apply-merge can validate and act:
```
{"id": "d1", "kind": "alias", "variant": "Taig", "candidate_source": "Tai", "choice": "yes_alias"}
```
Pipe the decisions JSON into apply-merge:
```
echo '{"auto_apply": [...], "decisions": [...], "consumed_chunk_ids": [...]}' \
  | python {baseDir}/scripts/merge_meta.py apply-merge "<temp_dir>"
```
Surface the summary JSON (auto_applied, decisions_resolved, consumed_chunks, errors) in your batch progress message.

apply-merge is transactional. If any decision is malformed (wrong choice for kind, missing fields, references a non-existent entity), the entire batch aborts with a non-zero exit and stderr details — no glossary mutation, no hashes recorded. On non-zero exit, fix the offending decision and re-pipe; prepare-merge will surface the same proposals because nothing was consumed.

Decision order in the input list is not significant. apply-merge internally dispatches entity-creating decisions before alias-attaching ones, so yes_alias decisions whose candidate is created by another decision in the same batch (a use_standalone_N, use_variant_N, or promote_to_separate_entity) succeed regardless of the order you pass them in. Alias chains (e.g. Taighi → Taig where Taig → Tai is also a pending alias decision) resolve via a fixed-point loop within the alias-attacher pass — you don't need to topo-sort or sequence chained aliases manually.

On a fresh run after a previous interrupted batch, prepare-merge will pick up any meta files left behind. Don't manually delete them.

4.6. Translate Table Captions and Headers

The chunk translation does not cover tables. A LaTeX table float is protected behind a ⟦T####⟧ placeholder so the math guard can guarantee its backslashes survive — and its \caption{} and column headers are inside that placeholder, so no translator ever sees them. The book comes out translated with its tables still in the source language, and every existing check passes, because they count tables, images and values and those are all correct.

Find out whether this paper is affected:

python {baseDir}/tests/format_probe.py "<temp_dir>" --lang <target_lang>

If it reports untranslated captions, run one sub-agent per file that holds tables — chunkNNNN.math.json for placeheld floats, output_chunkNNNN.md for inline ones — and have each translate:

the body of \caption{...} (and its optional [...] argument),
column header cells that are ordinary words (Method, Model, Bits, Budget, Avg., Total Time (s) …),
rotated or small-caps row-group labels,
prose inside \begin{tablenotes}.

and leave untouched: every number, every model/benchmark/method name, all LaTeX structure (& and \\ counts, \multicolumn/\multirow arguments, column specs, rules), and \cite{}/\label{} keys.

Tell each agent to edit with a Python script — json for the sidecars — never a shell heredoc.

Take the baseline before they start, and check it yourself afterwards. An agent that dropped an & or retyped a number reports exactly what a careful one reports, and nothing else in the pipeline would notice: the table is present, the caption is Korean, every count agrees, and the reader gets a number that was never in the paper.

python {baseDir}/scripts/verify_tables.py snapshot "<temp_dir>"     # BEFORE
#   ... the table sub-agents run ...
python {baseDir}/scripts/verify_tables.py check "<temp_dir>" --strict

The check requires the snapshot and refuses to answer without one, because a baseline reconstructed from the edited files is not a baseline. It allows any change to WORDS — captions, header cells, \textbf{} wrappers, notes — and fails on a changed number, a changed row count, a changed & count in any row, or a changed \multicolumn/\multirow span. Re-translate any table it rejects from the snapshot copy under <temp_dir>/.table_snapshot/.

An algorithm float needs the same pass, and used to need more. Its \caption{} is invisible to the translators for the same reason, and so is the prose inside \Comment{} and on the \Require/\Ensure lines — that is the pseudocode a reader actually reads. verify_tables.py cannot guard it: the float holds no tabular, so the snapshot records ZERO tables for a file that has an algorithm in it. Check the sidecar directly. The float itself now reaches the page — algorithm_float.py rewrites it as a numbered markdown list before pandoc sees it — but until this run every book silently lost it (KNOWLEDGE.md K78).

Then re-run the format probe; it must report zero untranslated captions.

A sidecar edit does not invalidate output.md — the merge is keyed on the output_chunk*.md mtimes, and --force-html only regenerates the HTML from the stale merge. Delete output.md before rebuilding, or the sidecar will read Korean while the PDF still shows English (KNOWLEDGE.md K42).

4.7. Tables and Equations Are Finished by the Build — Do Not Hand-Patch

Everything below happens inside merge_and_build.py on every run. It is described here so you recognise the output as intended and do not "fix" a book by editing HTML, and so you know what to look at if a paper comes out wrong.

Table captions get a number. float_units() numbers floats by counting \caption calls in flat.tex — calls, not environments, because one table* can hold two. number_table_captions() then prefixes each caption with 표 N (Table N) in the target language's form. A tabular takes the nearest caption above it, not the float's first one.

Table rules come from the source, not from a guess. table_structures() reads flat.tex and, per table in document order, records how many rows are header (\toprule … \midrule) and where the body rules fall: \midrule/\cmidrule hard, \addlinespace and \\[Npt] soft. apply_table_structure() promotes those header rows to <thead> (so they repeat on every page) and marks body rows rule-above / rule-above-soft. This runs whether the table reached the page as raw LaTeX or as a pandoc-converted markdown table. Only when the source has no rules at all does a heuristic add one above a trailing Average/평균 block.

Word repeats the header too. pandoc sets <w:tblHeader/> only when it found exactly one header row, so a two-deck header never repeated after a page break in Word. mark_docx_header_rows() marks the remaining rows from the same plan once pandoc has written the file.

References are coloured, and linked where the target is certain. link_cross_references() gives every figure, table and numbered equation an id, then turns 그림 N / 표 N / 식 (N) in the prose into anchors pointing at them. The colour is #001473, measured off the source preprint, which draws its own 332 links in exactly that. A caption is a LABEL, not a reference, so captions stay plain -- linking "그림 1 (Fig. 1)" to the figure it sits under is a link to itself.

Citations and 부록 A.8 get the colour but no link, deliberately. The bibliography arrives as raw \bibitem text with no per-entry anchor and the appendix subsections have no headings to anchor to, so a link would have to guess -- and a citation pointing at the wrong paper is worse than one that points nowhere. The pass runs on the body HTML AFTER tag_equations_in_html, because until then "식 (7)" has nothing to point at; the DOCX path never sees it and keeps plain text.

Grid tables are converted at merge time, not turned off. See KNOWLEDGE.md K66 — disabling grid_tables makes pandoc write the literal text [TABLE] and lose the table entirely.

Display equations are centred by display: flex; justify-content: center, not text-align, which Chromium ignores on a block <math>. The element stays full width so a right-margin (N) number still lands in the margin.

Check the result in the built HTML, not in the markdown:

python {baseDir}/tests/format_probe.py "<temp_dir>" --lang <target_lang> --strict

Every table must have a <thead>; the probe says how many do. A table with none means header_row_count() found no rule to stop at — read that table in flat.tex before touching anything else. A cell containing +--- or === means a grid table survived conversion: that is a bug in grid_tables_to_pipe(), not something to patch in the output.

5. Verify Completeness and Retry

After all batches complete, use Glob to check that every source chunk has a corresponding output file.

If any are missing, retry them — each missing chunk as its own sub-agent. Maximum 2 attempts per chunk (initial + 1 retry).

Then run the per-chunk gate over the whole book, not only over the last batch. A chunk can pass on its own and still be wrong about the book: a term the glossary gained in batch 3 was not in the table batch 1 was handed.

python {baseDir}/scripts/verify_chunk.py "<temp_dir>" --lang <target_lang> --strict --quiet

Every chunk must pass before you merge. A book built on a chunk that failed this gate cannot be repaired cheaply afterwards: re-converting moves every chunk boundary and renumbers every ⟦M####⟧, which discards the whole translation and every review of it.

6. Translate Book Title

Read config.txt from the temp directory to get the original_title field.

Translate the title to the target language. For Chinese, wrap in 书名号: 《translated_title》.

7. Post-process — Merge and Build

Run the build script with the translated title:

python {baseDir}/scripts/merge_and_build.py --temp-dir "<temp_dir>" --title "<translated_title>" --cleanup

If the user provided epub_cover, add --cover "<epub_cover>". If the user provided export_name, add --export-name "<export_name>".

The --cleanup flag removes intermediate files (chunks, input.html, etc.) after a fully successful build. If the user asked to keep intermediates, omit --cleanup. The chunk*.math.json sidecars are deliberately kept so a later re-merge can still restore formulas.

The script reads output_lang from config.txt automatically. Optional overrides: --lang, --author.

Other flags:

--build-only — skip the merge and build straight from an existing output.md. Needed for a temp dir already cleaned with --cleanup, where the chunks are gone.
--force-html — regenerate HTML even when book_doc.html looks newer than output.md.
--docx-engine {pandoc,calibre} — default pandoc, which produces native editable Word equations (OMML). Calibre's DOCX writer has no math support and drops every formula, so only use calibre if the pandoc output has a problem.
--pdf-engine {chromium,calibre} — default chromium, which renders through the local Chrome/Edge print engine and honours the @page print CSS, so page size and margins are what the profile says. Calibre honours neither and was silently producing US Letter pages with 25.4mm margins on every side; keep calibre only as a fallback when no browser is available.
--print-profile {a4-book,a4-large,a4-dense,letter-book} — default a4-book (A4, 18/18/22/18mm, 11.5pt body, 1.75 leading). Changing it requires --force-html, because HTML regeneration is keyed on output.md's mtime and the geometry is baked into book_doc.html at template time.
--math {mathml,none} — default mathml: offline math with no JavaScript and no external requests, which is also what survives into EPUB and PDF.
--allow-degraded-html — last resort. Without pandoc and without the markdown package the build now fails loudly instead of silently falling back to a regex converter that cannot render tables at all. Prefer installing pandoc or running python -m pip install markdown.

The build refuses to continue when a check fails, rather than shipping a quietly broken book: markdown or raw-LaTeX tables that did not become <table>, math present in output.md but absent from the HTML, image references that do not resolve, a math placeholder that a translator dropped, or a translated chunk whose LaTeX skeleton no longer matches its source.

Three things are repaired automatically on the way through, and reported:

Raw LaTeX tables. The arXiv backend keeps tables as \resizebox{..}{\begin{tabular}...}, often inside a \begin{table*} float. pandoc's markdown reader treats those as raw LaTeX blocks and drops them on the HTML path without warning, so each one is converted separately with pandoc -f latex -t html and spliced back in, caption and all.
Citations. A paper that ships a precompiled .bbl rather than a .bib has nothing for citeproc to read, so [@key] markers printed verbatim. They are numbered against the inlined \bibitem list. A key with no entry is left alone rather than given a wrong number.
Cross-references. (fig:x) / (tab:x) are resolved to Figure N / Table N (localised per language) using the \label order in flat.tex.

output.md itself is never rewritten by any of these — it stays the faithful merged translation. The resolved copy is written to prepared.md.

This produces in the temp directory:

output.md — merged translated markdown (placeholders restored to real LaTeX)
book.html — web version with floating TOC
book_doc.html — ebook version
book.docx — via pandoc, with editable equations
book.epub — via Calibre (requires Calibre)
book.pdf — via headless Chromium, laid out by the @page rule and the print stylesheet in template_ebook.html. Page numbers are stamped into the bottom margin by PyMuPDF afterwards, because Chromium implements no @page margin boxes. Falls back with --pdf-engine calibre.

7.5. Source-Fidelity Audit — dispatch a reader who holds both books

Run this on every paper, every time. It is not optional and it is not a formality. Every table and figure defect this pipeline has ever shipped was found by a person reading the finished page beside the original — never by a check. The probes count what they were taught to count; an omission is by definition the thing nobody counted.

Run tests/table_probe.py first: it is cheap and it catches what is countable — a lost \multicolumn span, a column that vanished, a value in the source tabular that is not on the page, a body row carrying numbers with no label. Fix what it finds before spending an agent.

Then dispatch one sub-agent per paper to do what no probe can. Give each:

the built PDF and the ORIGINAL PDF (ref_paper/<name>.pdf),
output.md, flat.tex, and the chunk*.math.json sidecars,
a scratch directory of its own, and filenames prefixed with the paper.

Tell it to render both pages and read them as images — PyMuPDF page.get_pixmap(matrix=fitz.Matrix(2,2)).save(path) — not to compare text dumps. Position is the whole point: a header that sits over the wrong column extracts identically to one that sits over the right column.

Ask for, per table and per algorithm float: the original page, the translated page, and either "matches" or a numbered list of concrete differences, each quoting the original cell text and the translated cell text. Then: column counts, row counts, group-header spans, row labels, footnote marks and tablenotes, superscripts still attached to their numbers, colour or bold that a caption refers to, and any cell squeezed until it cannot be read.

Require the agent to say what it could NOT verify. An audit that reports only findings is indistinguishable from one that stopped early.

Then check its claims yourself against the files. That is the point of asking for quotes and page numbers rather than a verdict.

7.5b. Then read the whole book, page by page

The audit above is scoped to tables and floats, and whatever it is not looking at goes unseen. A table-only pass signed off three books that were between them missing an entire 61-entry reference list, printing {=latex} in the body, carrying the 참고문헌 heading six pages inside its own list, and naming no authors on any title page. Every one of those was found by someone looking at pages.

Dispatch one sub-agent per paper to render EVERY page at fitz.Matrix(2, 2) and read the images — not text dumps; a heading sitting on top of a figure extracts perfectly. Ask for a line per page ("p7 — clean" or "p7 — "), then findings ranked by reader impact, then what it could NOT verify.

Run the probes alongside, plus a scan of the built PDF for markup that reached the reader: \command, $, :::, {=latex}, {-2em}, (tab:x), [@key], ??, HTML tags, <!--, and the ↩︎ of a footnote backlink. Those cost nothing and catch what an eye loses across ninety pages.

Keep adding to that list. Every entry on it is there because a build printed it and the scan of the day did not name it:  printed twenty-one times in AlphaQ while a scan looking for seven other shapes reported the book clean. A reader found it in a figure caption.

7.5c. Check the original before you change anything

A finding is a lead, not a verdict. Of the last four on one review's list, three needed no change: the headings carry no numbers because the PAPERS carry none, a unit prints italic because the source writes 96$tps$, and a rule reported missing was present in all 39 tables. Acting on the report alone would have produced three deviations dressed as repairs.

So for each finding, open the original and ask what it actually does. Change only where the book and the paper disagree. This is also how you catch your own regressions: it is what turned up 부록 A.10 printing as A.10절 after a fix made an hour earlier. See H23.

8. Report Results

Tell the user:

Where the output files are located
How many chunks were translated
The translated title
List generated output files with sizes
Any format generation failures

Report what the build actually said. If a check reported a shortfall — tables that did not convert, citations with no target, TOC entries whose page could not be resolved — say so with the numbers rather than rounding it up to "done".

9. Record What You Learned

Two logs, and the first decision is which one a finding belongs in:

the finding is…	it goes in	its number
a tool behaved in a way that surprised you	`KNOWLEDGE.md`	`K<n>`
your way of working cost you rework	`KNOWHOW.md`	`H<n>`

Why did this break? is knowledge. How should I proceed? is knowhow. One incident can produce an entry in each — the pandoc behaviour in one, the practice that would have caught it sooner in the other — but never the same entry twice.

Only if it is not already there. An empty diff is a fine outcome; most runs should produce one.

Add a KNOWLEDGE entry when you hit any of these:

An external tool (pandoc, Chromium, Calibre, PyMuPDF) behaved in a way that surprised you, and you have the measurement that proved it.
A fidelity check passed while the output was wrong, or failed while it was right.
You made a judgement call with numbers behind it (a size, a margin, a threshold) that the next run would otherwise re-derive.
Something turned out to depend on this machine — an installed font, a tool version, a path.

Follow the entry format and the rules in KNOWLEDGE.md's Maintenance protocol. Two of them matter most: add a row to the symptom index, because an entry nobody can find is not knowledge; and if you fixed the problem in code, write the test and compress the entry to one line pointing at it — the test is the record, the log keeps the why.

Add a KNOWHOW entry when the run taught you something about the METHOD:

You repeated a step you could have got right the first time, and can say what would have prevented it.
You found which document, check, or ordering has to come before which.
A practice paid off that is not obvious from the code — and you have the incident to prove it, not just the opinion.

Every KNOWHOW entry carries a *Cost when skipped:* line naming the real incident. A practice with no incident behind it is a preference, and preferences belong in AGENTS.md's conventions. Generic programming advice belongs in neither.

If a fix is worth enforcing rather than remembering, put it in code and tests instead. The verification gate is:

python -m compileall scripts tests
python -m unittest discover -s tests -p 'test_*.py'
python tests/layout_probe.py --strict
python tests/layout_probe.py --stress --strict
python tests/format_probe.py <temp_dir> --lang <lang> --strict
python tests/source_probe.py <temp_dir> --strict
python tests/table_probe.py <temp_dir> --lang <lang> --strict
python tests/inventory_probe.py <temp_dir> --lang <lang> --strict
python tests/leak_probe.py <temp_dir> --strict
python tests/consistency_probe.py <temp_dir> --lang <lang> --strict

Each looks at something the others cannot see, so none of them is redundant:

probe	question it answers
`layout_probe`	Is the page the size and type the profile asked for?
`format_probe`	Do DOCX, EPUB and PDF all carry the same tables and images?
`source_probe`	Do our numbers say what the original paper prints?
`table_probe`	Does each table still say what the source `tabular` says?
`inventory_probe`	Is any KIND of thing the paper contains missing from the page entirely?
`leak_probe`	Is anything on the page markup rather than words?
`consistency_probe`	Would a reader trip over the words?

leak_probe is the other half of that inversion. Scanning for {=latex}, {-2em}, ::: and the rest is a list of things that have already gone wrong, so it is always one build behind:  printed twenty-one times in AlphaQ while a scan looking for seven other shapes called the book clean, and a reader found it. This one asks what a sentence is made of instead -- Hangul, letters, digits, ordinary punctuation -- and reports every token carrying { } < > \ $ & | ^ ~. A lone one of those is rendered content (the bars of a norm, the braces of a set, the & between two authors); attached to anything else it is markup. The one other exception the probe grants is declared by the paper itself: a token the source spells with a LaTeX escape — R\&D, pick\&place, 50\% — is a word, because a column separator is written bare and only flat.tex still shows the difference. Name a new exception here when you meet a real one; do not go back to naming offences.

inventory_probe is the one that does not start from our own output. Every other check compares two artifacts this pipeline made, so it can only find a disagreement between stages — and it is blind, by construction, to what is absent from every stage at once. CafeQ shipped 61 references in output.md and none in the book while every count agreed with every other, because nothing counted references. Reading flat.tex and asking "the source has figures / a bibliography / an author — did any of it reach the page?" needs nobody to have thought of the check first. It reports only TOTAL absence: a count in LaTeX and a count in HTML do not line up, and treating that difference as a shortfall is how a check earns a reputation for crying wolf.

table_probe exists because nothing compared a table's CONTENT to the source. source_probe reads the original and counts NUMBERING — section, figure, equation, reference numbers — not one cell. verify_tables compares each table against a snapshot of OUR OWN files, so when a conversion drops a value before the snapshot is taken, the snapshot records the damage and the check passes forever. CafeQ shipped with twelve values missing from table 1 and six from table 5 while the prose went on citing them, and both checks were green. Every table defect in this project was found by a person reading the page.

source_probe is the one that compares against ground truth rather than against ourselves. Section numbers, figure numbers, equation numbers and cross-reference values are all reconstructed from flat.tex, and a reconstruction can be perfectly self-consistent and still wrong — an early version numbered SINQ's headings I. / A. / 1) because that is what IEEEtran does, while the paper prints 1. / 2.1.. Forty-one headings were labelled confidently and wrongly, and every other check was green.

format_probe is not optional. The layout probe measures the PDF and only the PDF, and twice that has let a DOCX regression through: raw HTML survives the HTML path alone, so a change that looks right in print can empty the Word file of every table or image without a single check complaining.

10. Repairing a book that is already translated

Sometimes a structural defect is only found after the prose is done and reviewed. Re-converting is the wrong instinct: it moves every chunk boundary and renumbers every ⟦M####⟧, so the planner asks to re-translate almost everything and the review is lost with it. Repair the artifacts in place.

Back up chunk*.md, output_chunk*.md and manifest.json first. Nothing here is under version control.
If the defect is in a raw LaTeX table, it never went near a translator — those are restored verbatim from chunk*.math.json at merge — so it can be fixed without touching prose at all:
```
python {baseDir}/scripts/repair.py sidecar "<temp_dir>"            # dry run
python {baseDir}/scripts/repair.py sidecar "<temp_dir>" --apply
```
This undoes the column spec pandoc emitted twice and lifts labels out of \rotatebox/\multirow. It is how SINQ went from ten ragged tables to none with nothing re-translated.
Otherwise patch the source chunk and the translated chunk together. The build comes from the output chunks, but the validators compare each output chunk against its source — images, math placeholders and LaTeX skeleton must still match, so an edit to one is an edit to both.
Re-hash and re-record every source chunk you touched:
```
python {baseDir}/scripts/repair.py rehash "<temp_dir>"
python {baseDir}/scripts/run_state.py record "<temp_dir>" chunk0005 ...
```
The source hash is how the pipeline decides a chunk needs translating again. Skip this and the next run silently re-translates a chunk that was already right — paying for it, and discarding the review.
Confirm nothing is queued before rebuilding: python {baseDir}/scripts/run_state.py plan "<temp_dir>" must report an empty translation_chunk_ids.
Touch the output chunks, then rebuild with --force-html. The merge is keyed on their mtime, so editing only a sidecar or a source chunk leaves output.md stale and the rebuild silently uses the old text.
Run the whole gate above.

The export aliases (--export-name) are written inside the temp dir. If you also keep copies somewhere the user reads, they are yours to refresh — a rebuild will not touch them, and a stale copy next to a fresh book.pdf looks exactly like a successful build.

In Ihrer bevorzugten KI fragen

Dokumentation