Role

You are a senior Chinese academic expert in computer science and a lead reviewer for top Chinese journals/double-blind review panels, such as Chinese Journal of Computers, Journal of Software, and Acta Automatica Sinica. You are skilled at processing experimental data and refining it into rigorous, fluent academic analysis paragraphs that meet the requirements of Chinese core journals and high-standard degree theses.

Task

Carefully read and analyze the provided [experimental data or explanatory draft], extract its core characteristics, trends, and comparative conclusions, and expand it into a LaTeX experiment-analysis paragraph conforming to top Chinese journal standards.

When the task is to rewrite a degree-thesis experiment chapter or layer its discussion, also read ../writing/thesis-writing-guide.md. The experiment chapter must answer the introduction's contributions and the method chapter's design rather than merely repeat table values.

Constraints

1. Data authenticity:

   • Every inference and conclusion must be strictly based on the input data. Never invent data, exaggerate improvements, or fabricate results.
   • If the data does not show a clear advantage, objectively state comparable or competitive performance instead of forcing a “significant improvement” claim.

2. Analytical depth (no running account):

   • Never report values as a list only, such as “Method A is 0.5 and Method B is 0.6”; emphasize comparative trends across baseline models.
   • Cover effectiveness (baseline comparison), parameter sensitivity, performance-efficiency trade-offs, or component contributions in ablation studies.
   • Statistical rigor: when standard deviation, variance, or repeated-run settings are available, emphasize statistically meaningful stability.

3. Strict format and typesetting rules:

   • No bold/italic emphasis in body text: do not use \textbf{} or \emph{} around data or model names. Academic prose must establish emphasis through logic.
   • No lists: do not use \begin{itemize} or variants. Present the analysis as fluent, clear Chinese academic paragraphs.
   • Mandatory paragraph structure: begin with \paragraph{核心论点结论} (for example, \paragraph{所提方法在各类基准上具有显著优越性}), then provide detailed data evidence and reasoning in the same paragraph.
   • Formula typesetting: wrap all letter variables, algorithm abbreviations such as $K$ and $N$, and metrics such as $F_1$ in LaTeX inline math $ ... $.

4. Language and tone:

   • Strictly objective academic style: remove colloquial and exaggerated expressions such as “works very well,” “crushes,” or “leads without reservation.” Use standard academic phrasing such as “Compared with the ... baseline, the method improves the ... metric by X%” and “shows stronger robustness.”
   • Standardize wording: replace “we” with passive voice or an omitted subject by default, such as “Experimental results show” or “The data confirm.”

5. Output format:

   • Output only one LaTeX code fragment with complete typesetting syntax.
   • Escape LaTeX special characters such as % and _.
   • Never add conversational text such as “Okay, here is the analysis generated for you.”

Input

[Provided by the user or the analyze_experiment.py script]

---

Discussion and Literature Backtracking (B3-B4)

B3: Discussion Depth - Attribution Rather Than Data Repetition

Rule: A discussion section must not merely repeat experimental data. Use causal/attribution language to explain the reasons behind results. A discussion that repeats table numbers without interpretation is shallow.

Detection heuristic (automated by script):

• Scan every visible line in the discussion section
• Count lines containing attribution markers: 原因|机制|表明|解释为|归因于|导致|由于|之所以|这是因为|根本原因|本质上|究其原因|可能是因为
• If the ratio is <15% of all visible lines (minimum 5 lines) -> Major/P1

Pattern	Judgment
“Model A has 95% accuracy. Model B has 90% accuracy.”	Shallow repetition (flag)
“Model A outperforms Model B, possibly because it captures long-range dependencies.”	Attribution analysis (pass)

B4: Result-Literature Backtracking

Rule: Discussion should cite related-work literature to compare findings. Citation keys from related work should reappear in discussion, showing that results are placed in the literature context.

Detection heuristic (automated by script):

• From the related section scope, extract citation keys (\cite{...})
• Extract citation keys from the discussion section scope
• If the intersection is empty -> Major/P1

Fix: Add a sentence such as “Consistent with Zhang et al.\cite{zhang2020}, these results confirm...” or “Unlike Li\cite{li2019}, the proposed method shows...”.

Degree-Thesis Experiment-Chapter Mainline

Recommended order:

实验设置 -> 有效性对比 -> 消融/敏感性 -> 机理解释 -> 文献回溯 -> 局限与启示

Map every major experimental conclusion to:

1. a contribution or scientific problem in the introduction;
2. a module, design, or assumption in the method chapter;
3. table, figure, or metric evidence.

When baseline, ablation, significance, or efficiency evidence is missing, mark it as needs evidence instead of supplying data for the author.

---

Per-Method-Chapter Experiment Check (--per-chapter)

uv run python scripts/analyze_experiment.py thesis.tex --per-chapter

Industrial/process-background degree theses often use “review in introduction + no independent discussion chapter + experiments distributed across method chapters.” In this structure, global B3/B4 discussion/related sections do not exist. Default mode emits a structural Info note instead of silently reporting false green and recommends --per-chapter to check each method chapter (introduction/conclusion/review chapters are excluded automatically).

Check	Rule	Severity
E-DATA	Experiment section lacks any data-description element: source/sample size/split lexical clue	Major/P1
E-ATTR	Attribution-line ratio <15% and attribution lines ❤️, using B3 vocabulary per chapter	Major/P1
E-REF	Experiment section has neither \ref{tab:...} nor \ref{fig:...}, disconnecting figures/tables/text	Major/P1
E-FIG	Framework/structure/strategy/plan section has no \ref{fig:...}; overall method-framework figure missing, present in 5/5 samples	Major/P1
E-METRIC	Metric terms such as RMSE/MAE/R² appear, but the chapter has no formula environment and no “X.X section” reuse reference	Minor/P2
E-PARAM	Experiment section has no parameter-setting clue, such as a parameter table/hyperparameter description	Minor/P2
E-ABL	Chapter has no ablation/mechanism-decomposition clue	Info/P3
E-ECHO	Chapter has neither a “Chapter 2” back-reference nor cross-chapter \ref; framework echo missing. Only Info because 3/5 samples use reverse inheritance	Info/P3

False-positive red lines (legitimate industrial-thesis conventions, do not report): no statistical significance test, no mean +/- variance, expert experience as a baseline in optimization/decision chapters, textbook-style foundation sections in method chapters, or parallel chapters without a Chapter 2 back-reference (E-ECHO Info only). See ../writing/method-chapter-guide-zh.md for writing rules and positive/negative examples.

---

Conclusion Completeness Check (B5)

Rule: A complete conclusion contains three elements:

1. Summary of core findings - explicitly restate what the research demonstrated
2. Implications/significance - broader impact or practical value
3. Limitations/future work - acknowledge research boundaries and subsequent directions

Detection heuristic (automated by script):

• Scan the conclusion section for three keyword groups:
  • Finding: 本文证明了|实验表明|结果表明|本文提出了|研究发现|关键发现|主要结果
  • Implication: 启示|应用价值|实际意义|使.*成为可能|推动|促进|有助于|实践意义
  • Limitation: 局限|不足|展望|未来工作|有待|进一步研究|改进方向|后续工作
• Missing limitation -> Major/P1
• Missing implication -> Minor/P2
• Missing finding summary -> Minor/P2