Module: Conclusion

Trigger: 结论, 总结与展望, 结论与展望, 展望, 结论章, conclusion, 结论检查, conclusion check

Commands

uv run python -B scripts/analyze_conclusion.py main.tex
uv run python -B scripts/analyze_conclusion.py main.tex --json

For a multi-file project, pass the main.tex entry point. The script assembles \input/\include files and locates the conclusion chapter (结论/结论与展望/总结与展望). Conclusion-to-Chinese-abstract comparison reuses the extracted abstract text.

Details

Perform in-chapter content-structure diagnosis plus conclusion-abstract comparison on the conclusion chapter. The 13 CC-* checks are detailed in the checker mapping and severity notes in ../writing/conclusion-guide-zh.md:

• Three-part structure (CC-TRIAD): summary body + innovation statement + outlook must all be present. Missing outlook/summary -> Error; missing innovation statement -> Warning.
• Opening inheritance (CC-OPEN), numbered contributions (CC-ENUM), and contribution skeleton (CC-SKELETON): opening ordinal phrases connect the research chain; list 3-4 contributions as (1)(2)(3); each follows the “for... propose... results show...” skeleton.
• Outlook (CC-OUTLOOK-EMPTY/TRANS/COUNT): empty-phrase blacklist + limitation-to-outlook transition + 2-3 items.
• Conclusion is not the abstract (CC-VERBATIM): sentence-level difflib comparison; verbatim repetition ratio >=30% -> Warning.
• Numeric consistency (CC-QUANT): every number in the conclusion must be found in the body; missing numbers produce a soft NEEDS-LLM prompt.
• Prohibitions (CC-NO-FIG / CC-NEW-CONCEPT): do not introduce new figures/tables in the conclusion (Error) or new concepts ([LLM]).
• Style Info (CC-RATIO / CC-SUBSEC): summary-to-outlook length ratio and subsection/chapter numbering style are informational only, not judged right or wrong.

Lane Separation

• [Script]: CC-TRIAD/OPEN/ENUM/OUTLOOK-*/VERBATIM/QUANT/NO-FIG/RATIO/SUBSEC are script-determined and output % CONCLUSION (源文件:L##) [Severity] [Priority] [Script]: CC-码 说明.
• [LLM]: CC-SKELETON (semantic completeness of the contribution skeleton) and CC-NEW-CONCEPT (new-concept detection) use the LLM lane. The script provides only prompts/coarse counts for agent judgment.

Boundaries with Other Modules (Do Not Duplicate Reports)

• \cite, maximum length (<=2000), vague wording -> use spec-check (check_spec.py). This script's footer routes there and does not repeat hard conclusion-format rules.
• Over-claiming (“first/fully solves/comprehensively outperforms”) -> use the ../writing/over-claim-guard.md workflow.
• English-abstract tense/AI tone -> use the deai module.

Skill-layer response:

1. Return findings as % CONCLUSION (L##) [Severity] [Priority], separating “check results” from “rewrite suggestions.”
2. For Errors (missing outlook/summary or figures/tables in the conclusion), give structural repair suggestions first.
3. For CC-OUTLOOK-EMPTY empty phrases, provide a rewrite example with a concrete technical direction.
4. Never fabricate results or add contributions not in the original text; leave \cite/\ref/\label/math environments unchanged by default.

See also: ../writing/conclusion-guide-zh.md, the dedicated conclusion-chapter writing guide with structural templates, positive/negative examples, and the checker mapping.