academic-paper-writer-pro

Academic Paper Expert (Main Router)

[!IMPORTANT] This file isMain router entry + Pipeline B/C definitions + typesetting format specification library. For detailed specifications of Pipeline A (OCR pipeline), please refer to ocr_kb/SKILL.md. DOCX generation for all pipelines uses docx/SKILL.md the methods defined in (docx-js to create a new document / unpack-edit-pack to edit an existing document).

0. Directory Convention

[!IMPORTANT] All intermediate files and generated outputs must strictly follow the directory structure below,placing any generated files in the project root directory is prohibited(including .md、.py、.docx , etc.).

项目根目录/
├── resources/
│   ├── pages/          # Pipeline A: 切分出的单页 PNG 图片
│   ├── figures/        # 所有管道: 配图（裁剪的或提取的）
│   ├── md/             # 所有管道: 提取/拆分结果
│   │   ├── page_N.md       # Pipeline A 按页
│   │   └── section_N.md    # Pipeline B/C 按章节
│   ├── scripts/        # 所有 Python 辅助脚本（可跨任务复用，清理时不删除）
│   ├── compiled_paper.md  # 最终汇总的完整 Markdown
│   ├── config.json     # 任务配置（源文件、格式、管道类型），整个任务期间不变
│   └── checkpoint.json # 运行时进度（当前单元、计数器、失败项），随处理进度更新
├── outputs/            # 最终交付的 .docx / .bib 文件 + 核查点中间版本
├── ocr_kb/             # Pipeline A: OCR 工作流 Skill
│   └── SKILL.md
├── docx/               # Word 文档操作 Skill（DOCX 生成的技术实现基础）
│   └── SKILL.md
├── pdf/                # PDF 操作 Skill
│   └── SKILL.md
├── content_generation/ # Pipeline D: 论文内容智能生成 Skill
│   └── SKILL.md
├── SKILL.md            # 本文件（主入口路由 + Pipeline B/C + 排版规范库）
└── <source>.*          # 用户提供的原始文件 (.pdf / .doc / .docx / .md / 项目代码目录)

1. New Task Startup Protocol (Step 0: Pre-flight)

[!CAUTION] The following four checks must be performedanybefore any writing or typesetting actionall completed, and cannot be skipped.

1.0 Skill Version Check & Auto Update

Each time processing is started, the user's local skill version must first be compared with the current latest version:

Run the corresponding command (such as git fetch origin && git status -uno) to detect the local repository status.
If the local version is inconsistent with the latest version: you should proactively prompt and help the user update automatically:

"Detected that the academic-paper-writer-pro skill has a latest version. Would you like help updating automatically? Reply update to upgrade, or skip to continue the current task."
If the user chooses to update, automatically execute the update command (such as git pull or npx skills add https://github.com/tfboy1/academic-paper-writer --skill academic-paper-writer-pro), and proceed to the next step after the update is complete.

1.1 Environment Cleanup Confirmation

Detect whether old files exist under resources/ in pages/、md/、figures/ as well as resources/checkpoint.json .

If old files exist: list the file inventory and explicitly ask the user:

"Detected intermediate files from the previous task (N pages/chapters, M Markdown files, K figures). Delete them to avoid interference? Reply delete to clear, or keep to continue the previous task."
If old files do not exist: skip this step.
When the user chooses "delete": delete resources/pages/*、resources/md/*、resources/figures/*、resources/checkpoint.json、resources/config.json。resources/scripts/ do not delete(scripts can be reused).
When the user chooses "keep": read resources/checkpoint.json, and enter the breakpoint recovery process (see §1.3).

1.2 Source File Confirmation and Pipeline Selection

Scan all supported format files in the project root directory (.pdf、.doc、.docx、.md）：

only 1: automatically selected, and the pipeline is determined based on the extension.
multiple: List all file names, types, and sizes, and require the user to explicitly specify the target file.
None: Error, require the user to place the source file.

Pipeline auto-selection rules：

Extension	Pipeline	Progress unit	Core Skill
`.pdf`	Pipeline A — OCR pipeline	page	`ocr_kb/SKILL.md`
`.doc` / `.docx`	Pipeline B — Re-layout pipeline	section	`docx/SKILL.md` + this file §3.2
`.md`	Pipeline C — MD direct conversion pipeline	section	`docx/SKILL.md` + this file §3.3
None / directory	Pipeline D — Content generation pipeline	section	`content_generation/SKILL.md`

1.3 Checkpoint recovery detection

Check whether resources/checkpoint.json exists and is valid:

exists and status == "suspended": Show the user the previous progress (number of completed units / total units, pipeline type), and ask:

"A previous task was detected (Pipeline X, N/M units completed). Continue from the checkpoint? Reply continue to start from unit N+1, or restart to clear all intermediate files."
does not exist or status == "completed": Start a new task normally.

2. File integrity check and resource preparation

2.1 File integrity check

Require the user to place the manuscript to be formatted and the formatting requirements in this directory:

Draft: The user's content file (.pdf、.doc、.docx or .md）。
Style Guide/Template：.docx template or .pdf guide (such as an IEEE template).
References: Ask the user whether they have a .bib file. If so,must be used first。

2.2 Format specification parsing (required for all pipelines)

[!IMPORTANT] Before splitting content or extracting text,format specification parsing must be completed first, to determine all layout parameters.

If the user provided a .docx template（config.template_file is non-empty):
- Use docx/SKILL.md 's unpack method to open the template
- Extract style definitions (heading levels, body font/font size, line spacing, margins, headers and footers, column settings)
- Record the extracted parameters in config.json 's format_params field
If the user specified a format name(such as "IEEE", "APA"):
- From this file's §6 Layout format specification library , read the corresponding preset parameters
- Record the parameters in config.json 's format_params field
If both are present: The template takes priority, with preset parameters as a fallback.

2.3 Resource directory preparation

Ensure that resources/(including subdirectories pages/、md/、figures/、scripts/) and outputs/ exist; create them automatically if they do not exist.

2.4 Task configuration persistence

Create or read resources/config.json(for detailed format, see ocr_kb/SKILL.md §0.5), record the source file path, file type, pipeline type, format specification, and typesetting parameters, so that subsequent continuations do not need to specify them again.

3. Core processing flow (dispatched by pipeline)

3.1 Pipeline A — OCR pipeline (PDF input)

[!IMPORTANT] All detailed specifications of Pipeline AAll detailed specificationsare defined in ocr_kb/SKILL.md . This section only provides a high-level description.

Image splitting: run resources/scripts/split_pdf.py to split the PDF into single-page PNGs.
Page-by-page loop: according to the specifications in ocr_kb/SKILL.md , extract page by page + append DOI to DOCX page by page.
Verification and suspension: verify every 2 pages, suspend every 4 pages.
Consolidate and finalize: merge all page_N.md → compiled_paper.md → final DOCX.

DOCX generation method: use docx/SKILL.md to create a new document with docx-js, and configure styles according to the parameters in the §6 format specification library.

3.2 Pipeline B — Re-layout pipeline (.doc / .docx input)

Applicable scenario: the user already has a Word document and needs lossless layout according to a new format specification (such as IEEE). To absolutely preserve the original native complex formulas (OMML), charts, and comments, this pipelinestrictly prohibitsthe use of a Markdown intermediate extraction scheme, and must use low-level XML appearance translation (Unpack-Edit-Pack).

Step 1: Format conversion (.doc only)

If the input is .doc(legacy format), first convert it to .docx：

python docx/scripts/office/soffice.py --headless --convert-to docx <input.doc>

Step 2: Unpack and extract specifications

Parse the target format parameters according to §2.2.
Extract the input DOCX to the native XML directory:

python docx/scripts/office/unpack.py <input.docx> resources/unpacked/

Step 3: XML-level style restructuring (Edit XML)

This step does not touch the logic of the main content. Write a local script to directly “reskin” from the low level:

Replace Styles (word/styles.xml): for the format specification (such as IEEE), forcibly replace all core styles in styles.xml with preset style names and fonts.
Force set <w:sectPr>: in word/document.xml , locate the section configuration block, and forcibly mount the two-column setting (such as <w:cols w:num="2" w:space="360"/>) and the required paper size (such as US Letter: <w:pgSz w:w="12240" w:h="15840"/>）。
Rebind <w:pStyle>: traverse all <w:p>, determine their hierarchy level by text features or the weight of the old style, and mount them as IEEE-standard Heading1、Normal. This ensures that the formula structures encapsulated in paragraphs at the same level are never touched <m:oMath>。

Step 4: Package and finalize in original form (Pack)

Seamlessly package and verify the reconstructed appearance:

python docx/scripts/office/pack.py resources/unpacked/ outputs/<name>_final_<date>.docx --original <input.docx>

This completely abandons the high-risk Markdown intermediate generation method, achieving 100% lossless preservation of content and layout!

3.3 Pipeline C — MD direct conversion pipeline (.md input)

[!CAUTION] It is absolutely forbidden to use Pandoc for one-click direct conversion from Markdown to Word! Pandoc will lose all highly customized formats required for Chinese degree theses or top conferences (such as three-line table rendering, strong first-line indent binding, etc.). It must and can only strictly call the docx/SKILL.md defined in docx-js (Node engine) or the low-level unpack -> edit XML -> pack script.

Applicable scenario: The user already has a paper in Markdown format and needs to convert it to a formatted Word document.

Step 1: Format specification parsing

Parse the target format specification according to §2.2 and determine all typesetting parameters.

Step 2: Chapter splitting

Read the .md file, split byLevel 1 heading（# ...) into multiple chapters.
Save each chapter as resources/md/section_N.md。
Count the total number of chapters.
Create resources/config.json（pipeline: "C", unit_type: "section"）。
Create resources/checkpoint.json（current_unit: 1, reset the counter to zero).

Step 3: High-definition pre-rendering and self-healing of Mermaid architecture diagrams (critical)

If the Markdown contains mermaid code blocks, before injecting into Word, an independent high-definition rendering pipeline must be executed:

Extract and save: Save the Mermaid code block as an independent resources/figures/fig_N.mmd file.
Mount environment interceptor: Limited by Windows' strict process penetration policy, it is necessary to dynamically generate in the root directory puppeteer-config.json(content:{"args": ["--no-sandbox", "--disable-setuid-sandbox"]}）。
Call CLI to compile: Execute the command npx mmdc -i resources/figures/fig_N.mmd -o resources/figures/fig_N.png -b transparent -p puppeteer-config.json。
Compilation fault-tolerant self-healing (Auto-healing): If mmdc throws an error (often caused by non-rigorous syntax such as spaces in subgraph names, brackets/single or double quotes in state scalars, etc.),The Agent must capture the error log and proactively use Python or regular expressions to fix .mmd incomplete syntax in the text, then retry the compilation command until the high-definition .png is successfully produced. It is strictly forbidden to silently skip or leave garbled characters at this step!
DOM placeholder replacement: Replace all complex Mermaid blocks in the original Markdown with standard image reference syntax ![Figure N](resources/figures/fig_N.png), so that they can be natively embedded in the next step.

Step 4: Chapter-by-chapter conversion and typesetting

Use docx/SKILL.md docx-js to create a new DOCX document (all styles must be mapped through code!), configure the styles according to the format parameters in §6, and then convert the content chapter by chapter:

Formula conversion：
- Inline formula $...$ → native OMML (via ocr_kb/scripts/latex_to_omml.py, see the interface in ocr_kb/SKILL.md §0.7）
- Display formula $$...$$ → OMML formula paragraph, with the number right-aligned
- Update checkpoint.json the global_equation_count
Table conversion：
- Markdown table → Word table XML (according to the table typesetting parameters in §5.3 + docx/SKILL.md Tables chapter,Force output of academic three-line tables）
- Update global_table_count
Image embedding：
- ![caption](path) → Check whether the image file exists
- Copy the image to resources/figures/(if not already there), globally number
- Embed into DOCX according to the image typesetting parameters in §5.2
- Update global_figure_count
Text typesetting：
- Heading level mapping (# → Heading 1，## → Heading 2，...）
- Apply font, font size, paragraph spacing, and alignment according to the format parameters in §6.In particular, precise control of the first-line paragraph indent (2 characters)。
Append to the output DOCX.
Update checkpoint.json。

Step 5: Verification and context protection

Exactly the same as Pipeline B (check every 2 chapters, suspend every 4 chapters).
Add physical interception: if it is detected that table data and sufficient page capacity were not generated as required, discard the generated docx.

Step 6: Consolidate and finalize

All section_N.md were created during splitting, merged into resources/compiled_paper.md。
The final DOCX is saved as outputs/<name>_final_<date>.docx。

3.4 Pipeline D — Content Generation Pipeline (Project Records/Code Input)

Applicable scenario: the user does not have a complete paper draft, only code, experimental data, or notes, and wants to intelligently generate an academic-format paper draft based on existing project materials.

For detailed operating specifications, see content_generation/SKILL.md, main process overview:

Analyze the various technical materials provided by the user, extract and confirm the overall outline structure of the paper with the user.
Build high-quality academic text by section-by-section autoregressive generation, saved to resources/md/section_N.md。
After generation is complete, the system automatically routes it to Pipeline C (MD direct conversion pipeline), reusing its typesetting capabilities to output the final style.

3.5 Common Rules for All Pipelines

The following rules apply regardless of pipeline and aremandatory：

Rule	Description
Process unit by unit	Whether it is a page or a chapter, each time a unit is completed, append it to the DOCX and update the checkpoint
Check every 2 units	Compare against the source file to confirm there is no information loss or formatting corruption
Suspend every 4 units	Save an intermediate version and notify the user to send "continue" to refresh the context
needs_review zero-clearance gate	Before final consolidation,`needs_review` must be empty
Retain intermediate versions	Checkpoint versions are not deleted, to allow rollback
Global numbering	The three counters figure / table / equation increment globally and are not reset across units

4. Reference Management

Priority: parse the .bib file provided by the user.
Supplement: if the user did not provide .bib，Do not fabricate or search for references yourself. Mark all missing citations (list the citation marker names) and notify the user to supplement them manually.
Output: generate the reference file in outputs/references.bib .

5. General Formatting Details (Cross-Pipeline Formatting Details)

[!IMPORTANT] The following details apply to all pipelines; specific parameters are overridden by the preset values in the §6 format specification library or by the user template. For the technical implementation of DOCX, uniformly refer to docx/SKILL.md。

5.1 Template File Usage Process

When config.template_file is non-empty:

Method 1 (recommended): use the unpack method of docx/SKILL.md to open the template, extract the style definitions in word/styles.xml , then copy these styles when creating a new document with docx-js.
Method 2: directly copy the template file as the output file, then replace the content using unpack → edit XML → pack, preserving all of the template's styles and page settings.

When config.template_file is empty:

Use docx-js to create a brand-new document, and manually configure styles according to the preset parameters in the §6 format specification library.

5.2 Image Layout Parameters

Parameter	Default Value	Description
Width Strategy	Scale proportionally to 80% of the column width	Single-column paper = page content width × 80%; double-column paper = column width × 95%
Position	Inline embedding, centered alignment	Do not use floating positioning
Caption Format	"Figure N: description text"	Below the figure, centered, 1pt smaller than the body text, bold the "Figure N" part
Spacing Between Caption and Body Text	6pt above, 12pt below
Image Quality	Original resolution, no compression

docx-js Implementation Reference：docx/SKILL.md Images section (§220-232), note that type parameter is required,altText all three subfields are required.

5.3 Table Layout Parameters

Parameter	Default Value	Description
Width	Full column width (DXA units)	Single-column = page content width; double-column = column width
Border	Thin line `#CCCCCC`，1pt
Header	Bold, light blue background `#D5E8F0`	Use `ShadingType.CLEAR`(do not use SOLID)
Cell Padding	80 DXA top and bottom, 120 DXA left and right
Table Note Format	"Table N: description text"	Above the table, centered, 1pt smaller than the body text

docx-js Implementation Reference：docx/SKILL.md Tables section (§173-219), must set both columnWidths and each cell's width, and must use WidthType.DXA。

5.4 Headers, Footers, and Page Numbers

Parameter	Default Value	Description
Page Number	Bottom center, Arabic numerals
Header	Empty (unless required by the format specification)	IEEE has no header, theses usually do
Special First Page	No header or page number on the first page (if required by the format)

docx-js Implementation Reference：docx/SKILL.md Headers/Footers section (§251-268), use PageNumber.CURRENT。

5.5 Column Layout

Parameter	Default Value	Description
Number of Columns	Determined by the format specification (IEEE = 2 columns, APA = 1 column)
Column Spacing	360 DXA（0.25 inch）
Title/Abstract	Usually spans columns (separated by section break)

docx-js Implementation: through the sections property, set different sections to have different numbers of columns:

// 标题区域 - 单栏
{ properties: { column: { count: 1 } }, children: [/* 标题、作者、摘要 */] },
// 正文区域 - 双栏
{ properties: { column: { count: 2, space: 360 } }, children: [/* 正文 */] }

5.6 Formula Layout

Parameter	Default Value	Description
Inline Formula	OMML `<m:oMath>`, on the same line as the body text
Display Formula	OMML `<m:oMathPara>`, centered
Formula Number	Right-aligned, format `(N)`
LaTeX → OMML	via `ocr_kb/scripts/latex_to_omml.py`

docx-js / XML implementation reference：docx/SKILL.md Formulas (OMML) section (§447-469).

6. Format Style Library

[!IMPORTANT] Format presets are stored as separate .md files in the templates/ directory, and each file contains a complete table of formatting parameters. When the user specifies a format name, read the corresponding file. If the user provides a .docx template, the settings in the templateoverridethe presets.

6.1 Built-in Format Presets

Format Name	File	Applicable Scenarios
IEEE	`templates/ieee.md`	IEEE conference papers, journal papers
APA 7th Edition	`templates/apa7.md`	Psychology, social science papers
Chinese degree theses	`templates/chinese_thesis.md`	Domestic university bachelor's, master's, and doctoral degree theses
ACM	`templates/acm.md`	ACM computer society conference and journal papers
Springer LNCS	`templates/springer_lncs.md`	Academic monographs such as computer science lecture notes
NeurIPS	`templates/nips.md`	NeurIPS and machine learning-related conferences
MLA	`templates/mla.md`	Papers in the humanities and language and literature fields

Lookup Rules：

User says "format according to IEEE" → read templates/ieee.md
User says "format according to APA" → read templates/apa7.md
User says "format according to degree thesis" / "format according to graduation thesis" → read templates/chinese_thesis.md
User says "format according to ACM" → read templates/acm.md
User says "format according to Springer" / "format according to LNCS" → read templates/springer_lncs.md
User says "format according to NeurIPS" / "format according to NIPS" → read templates/nips.md
User says "format according to MLA" → read templates/mla.md
No match → enter the §6.2 user-defined process

6.2 User-Defined Format

When the format specified by the user is not among the built-in presets:

Ask the user to provide a .docx template or explicitly describe the formatting requirements.
Extract parameters from the template or description and fill in config.json 's format_params。
If the user's description is incomplete, read templates/ieee.md as fallback default values for missing parameters.

6.3 Extending New Formats

If you need to add a new format preset (such as ACM, Springer LNCS, etc.), simply create a corresponding new templates/ in the .md directory and fill in the parameters according to the table structure of the existing template.

7. Summarize Output After Completion

After formatting is complete,mustoutput the following structured completion report:

【完成报告】
源文件：<config.source_file>
输入类型：<config.source_type>
处理管道：Pipeline <config.pipeline>
格式规范：<config.format_style>
总单元数：<checkpoint.total_units>（<config.unit_type>）
提取公式数：<checkpoint.global_equation_count>
裁剪/嵌入配图数：<checkpoint.global_figure_count>
提取/转换表格数：<checkpoint.global_table_count>
核查点版本数：<outputs/ 中 checkpoint 文件的数量>
最终文件：outputs/<最终文件名>.docx
遗留问题：<needs_review 中的单元编号 / 无>

Then proactively confirm the following optional value-added services with the user:

Logical verification: check the argumentative logic between paragraphs.

8. Troubleshooting & Best Practices for Formatting and Environment

[!WARNING] When performing automatic formatting and document generation, Markdown parsing and underlying environment conflicts are often encountered. The following preventive specifications must be followed in the formatting pipeline:

8.1 Regex Stripping for Markdown Nested Bold Parsing

Problem：marked.lexer When processing complex nested paragraphs (such as bold **text**inside lists), because Inline-level tokensis not deeply recursively extracted by default, it is easy for ** symbols to leak directly into the DOCX output as plain punctuation text. Specification: when constructing docx-js When the Paragraph text is mounted,completely abandon AST-based style markup matching, and uniformly switch to prefix custom regular expressions (e.g. text.split(/(\*\*[\s\S]*?\*\*)/g)) for explicit splitting, mapping them directly to new TextRun({ text: innerText, bold: true })。

8.2 Extreme Fragility of the Mermaid Compiler and Syntax Cleaning

Problem：npx mmdc : npx mmdc is extremely picky about generated code. For example, usecaseDiagram , is deprecated and unsupported in the new version, and in erDiagram , if an entity attribute is followed by an outside comment with a double quote " , it will directly cause the rendering thread to deadlock and the command line to crash. Standard: Before invoking the execution environment, use a replacement script to clean non-compliant keywords (forcibly convert the old syntax to flowchart), and strictly remove dangerous quotes. To prevent Windows process isolation errors, the configuration file parameters need to mount {"args": ["--no-sandbox", "--disable-setuid-sandbox"]}。

8.3 Cross-platform Regular Expression Newline Character Trap (\r\n)

Problem: When using Regex for cross-platform operations to forcibly extract boundary-closed code blocks (e.g. /```mermaid\n([\s\S]*?)```/g), once a file written by a Windows host uses \r\n carriage return and line feed characters, the match will directly fail, causing large areas of lost images. Standard: Whenever multi-line text cutting is involved, be sure to introduce the fault-tolerant wildcard \r?\n as the boundary detector, and never hard-code a single newline type.

8.4 Crash from Implicit Required Items of the DOCX-JS ImageRun Component

Problem: When calling ImageRun to insert a Buffer image stream engine, if the altText parameter nesting combination is omitted, the compilation pipeline will silently crash with no clear stack feedback (Packer throw catch unclear). Standard：ImageRun : When initializing the object, be sure to fully complete and mount the three indispensable security metadata fields:altText: { title: "x", description: "x", name: "x" }, as well as statically specifying type: "png"。

8.5 Protection Against Operation Handle Lock (EBUSY) During Automated Updates

Problem：fs.writeFileSync : When intending to output and refresh the target document, if the document happens to be occupied within the user's browsing window process, it triggers EBUSY: resource busy or locked and blocks exit. Standard: The export-side name of the automation script should dynamically concatenate an auto-incrementing numeric suffix or a dynamically flowing timestamp (e.g. _V3_CLEAN), so as to avoid the conflict disaster of forced overwrite locks at the system level.

9. Academic Integrity Gate — 7-Mode Anti-Hallucination Blocking

[!CAUTION] The mechanism in this section applies to Pipeline D (Content Generation Pipeline) as a mandatory interception point after content writing is completed and before layout assembly. It cannot be skipped, and the Agent cannot decide on its own that "execution is unnecessary."

9.1 Trigger Timing

Pre-Review Gate: Triggered immediately after all chapter drafts section_N.md have been generated.
Final Gate: Triggered again after self-review revisions and before handing over to Pipeline C for layout.

9.2 7-Mode Blocking Checklist

The Agent must self-check item by item and output a structured report, marking each item as ✅ 通过 or ❌ 阻断（附具体位置与原因）：

#	Failure Mode	Check Content	Chinese Explanation
1	Disguising Implementation Flaws as Innovation	Whether bugs or omissions are packaged as "design features" during code parsing	Strictly prohibit glorifying code errors as "innovations"
2	Fabrication of experimental data	Whether the performance data in the testing chapter has actual code or log sources	Strictly prohibit fabricating figures such as latency and throughput out of thin air
3	Methodology drift	Whether the architecture described in the design chapter is consistent with the actual code structure	Strictly prohibit describing one thing while the code does another
4	Citation hallucination	Whether the references contain fabricated authors, years, or journal names	Mark all unverifiable citations as `[CITE_UNVERIFIED]`
5	Shortcut dependency	Whether calling a third-party API is directly equated with "this system implements algorithm XX"	Strictly prohibit confusing calling with self-development
6	Empty-shell chapters	Whether there are extremely short chapters of fewer than 300 words or a large number of `[TODO]` placeholders	Must be reworked and expanded
7	Repetitive discussion	Whether the same concept is repeatedly explained across multiple chapters (e.g., reintroducing front-end/back-end separation in every chapter)	Deduplicate and cross-reference

9.3 Blocking logic

If any one item exists ❌：Strictly prohibithand the document over to typesetting. The Agent must locate the problematic chapters and automatically fix them, then rerun this checklist.
All ✅: generate resources/integrity_report.md and continue the process.

10. Multi-Perspective Self-Review Panel

[!IMPORTANT] This section requires the Agent, after completing the first draft, toswitch to a reviewer personato conduct a multi-dimensional quality assessment. This mechanism is inspired by the peer review system, with the aim of maximizing the exposure of logical breaks and weak points in the argumentation in the absence of human reviewers.

10.1 Composition of the review panel (5 perspectives + devil's advocate)

The Agent must read the full text in sequence in the following 6 roles, with each role outputting an independent brief comment (in Chinese):

Role	Focus dimensions	Chinese description
Structure reviewer	Chapter logical chain, progression of argumentation	Is the outline clear? Are there logical breaks between chapters?
Technical reviewer	Accuracy of algorithm descriptions, reasonableness of system architecture	Do the technical details withstand scrutiny? Are there obvious engineering fallacies?
Language reviewer	Academic wording, sentence diversity, residual colloquialism	Are there repeated "de de de", colloquial expressions, or traces of AI formulaic writing?
Data reviewer	Completeness of figure/table citations, self-consistency of data	Do the figure/table numbers cited in the text actually exist? Do the data contradict each other?
Format reviewer	Heading levels, citation format, paragraph indentation	Does it conform to the typesetting requirements of the target academic template?
Devil's advocate	Specifically rebuts the core arguments	Try to attack the paper's weakest argument and raise the harshest doubts

10.2 Quantitative scoring table (0-100 Quality Rubric)

Dimension	Weight	Scoring Criteria
Argumentation Logic (Argumentation)	25%	Completeness of the reasoning chain and strength of causal links
Technical Depth (Technical Depth)	25%	Whether it truly reaches the core of the code rather than surface-level description
Language Quality (Language Quality)	20%	Avoid AI clichés, varied sentence structures, precise terminology
Media Richness (Media Richness)	15%	At least 1 figure and 1 table per chapter
Format Compliance (Format Compliance)	15%	Strictly align with the target template requirements

10.3 Decision Mapping

≥ 80 points: Pass, proceed directly to the typesetting pipeline.
65-79 points: Minor Revision; the Agent automatically revises the issues identified in the review and then re-scores.
50-64 points: Major Revision; the Agent must deeply rewrite the relevant chapters.
< 50 points: Reject; output a diagnostic report and notify the user to intervene.

10.4 Devil's Advocate — Anti-Sycophancy Protocol (Anti-Sycophancy Protocol)

After the Devil's Advocate raises an attacking challenge,it is not allowed to concede twice consecutively in a single adversarial exchange。
Before each "concession," the validity of the rebuttal must be evaluated on a 1-5 scale, and a concession may be made only when the evaluation is ≥ 4.
Set up a concession rate tracker: when the global concession rate exceeds 60%, trigger a "Frame-Lock Detection" alert, requiring the Agent to step back and examine the underlying assumptions.

11. Anti-Leakage & Context-Rot Prevention (Anti-Leakage & Context-Rot Prevention)

11.1 Knowledge Isolation Directive (Knowledge Isolation Directive)

Throughout the entire content generation process of Pipeline D, the Agent must abide by the following iron rules:

IRON RULE: The project code and materials provided by the user are the only Source of Truth. The Agent is prohibited from using "memories" from its own training data to fill any information gaps. If a technical detail cannot be substantiated in the project code, it must be marked [素材缺口] and that passage paused, rather than completed based on imagination.

11.2 Mid-Conversation Reinforcement (Mid-Conversation Reinforcement)

In multi-round chapter generation, whenever the Agent begins writing a new chapter (that is, switching from section_N.md to section_N+1.md), the following two reminders must be re-injected into the internal System Prompt:

Iron rule reiteration: "You are writing an academic paper. Your only source of truth is the project materials provided by the user. The use of memory filling is prohibited."
Anti-pattern detection: "Check whether what you are about to write contains large passages of repeated discussion with previous chapters. If so, replace the repeated passages with cross-references (such as 'as stated in Chapter X')."

11.3 Material Gap Protocol (Material Gap Protocol)

When the Agent discovers during writing that a technical detail is needed but was not provided by the user:

Prohibitedto silently fill it in using the AI's internal knowledge.
Mustinsert a marker at that location:[素材缺口：此处需要 XX 的具体实现代码/数据/配置]
List all material gaps uniformly in the final integrity report (§9), and notify the user to supplement them manually.

12. Writing Quality Check (Writing Quality Check) — Chinese Academic Specialization

[!IMPORTANT] This checklist is enforced during the "Language Reviewer" step of the §10 self-review assessment, and is used to detect common patterned traces in AI-generated text.

12.1 AI High-Frequency Word Alerts (25 High-Frequency Chinese Words)

The following words trigger a warning when they appear in a single paper more than the specified threshold:

Category	High-Frequency Word Examples	Per-Paper Limit
Empty Modifiers	effectively, significantly, greatly, fully, deeply	Each word ≤ 5 times
Clichéd Openings	With the development of..., as is well known, in recent years, it is worth noting that	Each word ≤ 3 times
Overpromising	perfectly, thoroughly, without a doubt, indisputably	Each word ≤ 2 times
Typical AI Words	empower, closed loop, lever, pain point, precipitate, align, granularity	Each word ≤ 2 times
Piled-Up Conjunctions	therefore, however, in addition, at the same time, in summary	Each type ≤ 8 times

12.2 Sentence Pattern Diversity Detection (Burstiness Check)

For 3 consecutive paragraphs, thefirst sentencemust not use the same sentence structure (e.g., three consecutive paragraphs all beginning with "This system...").
The standard deviation of paragraph length should be > 30 characters (to avoid machine-like uniform paragraphs).

12.3 Throat-Clearing Detection

Detect and delete uninformative opening remarks, such as "To better understand..." and "In this section, we will...".
Academic writing should enter directly into the core of the discussion, rather than previewing what is about to be said.

12.4 Punctuation Pattern Control

Em dashes (——) used ≤ 5 times per paper.
Exclamation marks (!) are absolutely prohibited in the main text (allowed only in non-main-text parts such as acknowledgments).

13. Revision Score Trajectory Tracking (Score Trajectory Tracking)

When a paper goes through multiple rounds of the "write → review → revise" cycle, the Agent must record the score of each round to prevent certain dimensions from unexpectedly regressing during revision:

| 轮次 | 论证逻辑 | 技术深度 | 语言质量 | 图表充实度 | 格式规范 | 总分 |
|------|----------|----------|----------|------------|----------|------|
| R1   | 72       | 68       | 75       | 60         | 80       | 71   |
| R2   | 78       | 75       | 73 ⚠️↓   | 70         | 82       | 76   |
| R3   | 82       | 80       | 80       | 75         | 85       | 81 ✅|

When any dimension's score compared with the previous rounddrops ≥ 5 pointsmark ⚠️↓ and explain the reason for the regression in the revision report.
Only when the total score reaches ≥ 80 points may it be handed over to the Pipeline C typesetting engine.