In-depth guide to logical connection and methodology

Paragraph-level logical connection (AXES model)

Components	Description	Examples
Assertion (claim)	Clear topic sentence, stating the core point of view	"The attention mechanism can improve the effect of sequence modeling."
Xample (example)	Specific evidence or data to support the claim	"In experiments, the attention mechanism achieved 95% accuracy."
Explanation (explanation)	Analyze why the evidence supports the claim	"This improvement stems from its ability to capture long-range dependencies."
Significance	Connection to a broader argument or next paragraph	"This finding informs the structure of this article."

Transition signal words

Relationship Type	Chinese Signal Word	English Correspondence
Progressive	In addition, further, more importantly	furthermore, moreover
Turning	However, however, contrary	however, nevertheless
cause and effect	therefore, hence, therefore	therefore, consequently
order	first, subsequently, finally	first, subsequently, finally
Example	for example, specifically, especially	for instance, specifically

Methodology in-depth checklist

• [ ] Each claim is supported by evidence (data, quotes, or logical reasoning)
• [ ] There is a good reason for the choice of method (why was this method chosen instead of another?)
• [ ] Explicit acknowledgment of study limitations
• [ ] Clearly state assumptions
• [ ] Reproducibility details are sufficient (parameters, data sets, evaluation metrics)

FAQ

Problem type	Manifestation	Fix method
Logical gaps	Lack of cohesion between paragraphs	Add transitional sentences to illustrate the relationship between paragraphs
Unsubstantiated claims	Assertions lacking supporting evidence	Supplementary quotes, data or reasoning
Shallow methodology	"This article uses X" but no reason	Explain why X is suitable for this problem
Implicit assumptions	Unstated preconditions	Explicitly stated assumptions

Chapter Guide

Chapters	Key points of logical connection	Key points of methodological depth
Introduction	Background → bottleneck/blank → scientific problem → smooth connection of contribution	Demonstrate research significance and defensible questions
Related work	Grouped by topic, explicit comparison	Positioning the relationship with previous work
Methods	Progression between chapters and modules	Demonstrate the motivation, design and technical advantages of each module
Experiment	Setup→Results→Analysis→Discussion→Limitations process	Explain the selection of evaluation indicators and the result mechanism
Discussion	The connection between discovery → revelation → limitation	Acknowledging research boundaries

When a chapter-level rewriting plan is needed, read thesis-writing-guide.md as a supplement, and take "research background -> technical bottlenecks/research gaps -> scientific issues -> methods of this article/chapter work -> experimental evidence -> contribution closure -> limitations and prospects" as the main line.

Output format

% 逻辑衔接（第45行）[Severity: Major] [Priority: P1]: 段落间逻辑断层
% 问题：从问题描述直接跳转到解决方案，缺乏过渡
% 原文：数据存在噪声。本文提出一种滤波方法。
% 修改后：数据存在噪声，这对后续分析造成干扰。因此，本文提出一种滤波方法以解决该问题。
% 理由：添加因果过渡，连接问题与解决方案

% 方法论深度（第78行）[Severity: Major] [Priority: P1]: 方法选择缺乏论证
% 问题：方法选择未说明理由
% 原文：本文采用ResNet作为骨干网络。
% 修改后：本文采用ResNet作为骨干网络，其残差连接结构能有效缓解梯度消失问题，且在特征提取任务中表现优异。
% 理由：用技术原理论证架构选择

Best Practices

Reference Elsevier, Proof-Reading-Service:

1. One paragraph, one topic: Each paragraph focuses on a single core idea
2. Topic sentence first: The beginning of the paragraph states the proposition of this paragraph.
3. Complete evidence chain: Each claim needs support (data, references or logic)
4. Explicit Transition: Use signal words to indicate paragraph relationships
5. Argument rather than description: Explain "why" rather than just state "what"

Completeness of the blurb after the title (S1)

Rule: Except for special sections such as the abstract, table of contents, references, acknowledgments, and appendices, each chapter, each section, each subsection, and the four-level headings responsible for the function of developing arguments should first give an introduction after the title, and then enter the list, chart, formula, or next-level heading.

Minimum functions of the introduction:

• Indicate what will be discussed at this level;
• Explain why it is discussed here;
• Explain the relationship with the above;
• Preview the unfolding sequence within this level.

Detection Heuristics (Script Automation):

• Scan \chapter, \section, \subsection, \subsubsection, \paragraph and other header commands;
• If the first thing that appears after the title is a list, chart, formula environment, next-level title, or there is no visible text, it is determined that the introduction is missing;
• If only a very short sentence appears after the title and there are no introductory signals such as "this chapter/this section/below/first", then marking it as an introduction may be too weak.

Revision suggestions: Add a complete introduction after the title to define the research object, writing purpose and writing route at this level, instead of directly piling up the content.

---

Literature review quality verification (A1-A4)

Make sure relevant work sections are comprehensive analyzes rather than simple lists.

A1: Topic clustering (author/year listing is prohibited)

Rule: Related work should be organized by research topic rather than by author or year of publication. Three or more consecutive sentences that follow the pattern "The author (year) proposed/introduced..." are regarded as lists.

Detection Heuristics (Script Automation):

• Regular: ^.*?[（(]\d{4}[)）].*?(?:提出|引入|设计|开发|采用|构建|建立)
• Threshold: 3 or more consecutive matches → Major/P1

Mode	Judgment
"Zhang San (2019) proposed X. Li Si (2020) introduced Y. Wang Wu (2021) designed Z."	Listing mode (mark)
"Attention-based methods have experienced continuous evolution... Zhang San (2019) and Li Si (2020) explored it from different angles... However, Wang Wu (2021) pointed out..."	Topic Synthesis (Passed)

Correction: Reorganized into thematic groups and conducted critical comparative analysis within the groups.

A2: Theme groups are followed by critical analysis (LLM judgment)

Rule: Each topic group should be followed by a comprehensive review that compares or evaluates the cited literature - not just lists. Look for evaluative language: "however," "despite," "common shortcomings," "in comparison."

*This rule requires LLM judgment, and regularity cannot be reliably detected. *

A3: Derivation of research gaps at the end of related work

Rule: The last paragraph of related work must contain a clear description of the research gap that provides motivation for this study.

Detection Heuristics (Script Automation):

• Scan the last 10 lines of chapter related
• Keywords: 研究空白|不足|然而.*?尚未|仍然.*?(?:挑战|困难)|有待|缺乏|尚未解决|亟待|亟需|鲜有研究|未能充分
• No match → Major/P1

A4: Funnel type citation density (LLM judgment)

Rule: The citation density should be in a funnel shape of wide→narrow→fine: first introduce the general field, gradually narrow down to specific sub-problems, and finally focus on the most relevant previous work. A flat or inverted funnel suggests poor narrative structure.

*This rule requires LLM judgment to evaluate narrative arcs. *

---

Cross-chapter logic chain closure check (C3)

Rule: Statements of contribution in the introduction must be clearly echoed in the conclusion. For example, if it is stated in the introduction that "this article proposes

Detection heuristic (script automation, full document mode is enabled by default; --cross-section is the compatibility switch):

• Extract contribution keywords from introduction chapter
• Extract response keywords from conclusion chapter
• If the introduction has a contribution statement but the conclusion has no responsive language → Major/P1 (observation marked [Script])

Introductory Statement	Desired Concluding Response
"This article proposes a novel attention mechanism."	"Experiments show that the proposed attention mechanism achieves..."
"The main contributions of this article are: (1)..."	"The experiment verified the contribution (1)..."

NOTE: This check is heuristic in nature and the results are presented as observations, not deterministic judgments.

---

Motor red line closure diagnosis (optional switch: --motivation-thread)

Rule: A strong dissertation follows a single "problem → solution" thread. Every promise made in the introduction ("This thesis proposes X to solve Y") should be validated in the experiments or results and answered in the discussion or conclusion. This diagnostic exposes breaks in that thread.

How to run (Additional type - without this switch, the logic module output is exactly the same as the original):

uv run python scripts/analyze_logic.py main.tex --motivation-thread

Output (read only, all marked [Script]):

• Commitment Mapping: Each promise in the introduction (hit CONTRIBUTION_KEYWORDS_ZH) → the line with the highest word overlap in the experiment/results; [未找到证据] means that the promise has never been verified.
• Closed mapping: Each claim in the introduction → the line with the highest word overlap in the discussion/conclusion; [未闭合] means that the claim has never been answered.
• Wandering evidence: A result line that carries a numerical conclusion but cannot be traced back to any of the introductory promises (possibly scope creep).

Mechanism: keyword + content token overlap (English word ≥4 characters + Chinese character tuple). It is a navigational aid rather than a conclusive statement—reports explicitly indicate possible false positives and require human review. If a promise is shown as unmatched simply because the results section has been worded differently, it should be regarded as a prompt to "make up for an explicit response" rather than a hard error.

Mode	Judgment
Introduction "This article proposes sparse attention to reduce latency" → Result "Latency reduced by 42%"	Matched (passed)
Introduction promises interpretability, but no results section mentions	[未找到证据] (tag)
Introduction claims never echoed in discussion/conclusion	[未闭合] (tag)

---

Chapter writing main line reinforcement

When a user requests to "rewrite the introduction/method chapter/experimental discussion/summary and outlook", first determine whether this is a main writing problem or a script diagnosis problem. If you need to write a plan, keep the script diagnosis results and output as follows:

1. Chapter objective: The scientific question or technical question to be answered in this chapter.
2. Paragraph roles: background, bottlenecks, gaps, methods, strengths, evidence, limitations, prospects.
3. Evidence mapping: Each contribution or innovation point corresponds to the methods chapter, experimental results or conclusion response.
4. Rewrite blueprint: Rewrite at the paragraph level only when the user explicitly requests it, and no new references, data or experimental conclusions will be added.