De-AI Module Reference

Purpose: Detect and reduce AI-generated writing traces while preserving LaTeX syntax and technical accuracy.

Core Principles

1. Syntax preservation: Never modify \cite{}, \ref{}, \label{}, math, or LaTeX commands
2. Zero fabrication: Never add data, metrics, comparisons, or claims
3. Information density: Every sentence must convey verifiable information
4. Restrained wording: Avoid unsupported certainty; use appropriate hedging
5. Academic payload first: Preserve the problem-method-evidence-conclusion-boundary chain before changing tone

Academic Humanization Contract

Before removing "AI flavor", extract four protected buckets:

• Facts/evidence: data, formulas, experiments, figures, tables, citations, and metrics.
• Claims/stance: the author's actual conclusion, method choice, uncertainty, and limitation.
• Logic: chapter role, paragraph role, claim-evidence mapping, and cross-chapter closure.
• Boundaries: scope, assumptions, unverified points, and 待补证 items.

Only after these are clear should the module remove structure shells. The default output remains diagnostic findings, a risk summary, or a rewrite blueprint. Provide prose rewrites only when the user explicitly asks for them.

High-Priority AI Patterns (Must Fix)

Pattern	Example	Fix
Empty adjectives	"显著提升"	Replace with specific metric: "MAE 降低 12%"
Absolute assertions	"显而易见", "必然"	Add qualification: "实验结果表明"
Vague quantifiers	"大量研究"	Use numbers: "三项研究 [1-3]"
Template openings	"近年来", "随着科技的飞速发展"	Start from specific problem context
Stacked citations	"[1]-[5]" without discussion	Discuss each cited work individually
Filler connectors	"总之", "不可否认的是", "值得注意的是"	Delete; state conclusion directly
Structure shells	"不是 A，而是 B", "真正的问题", "我的结论是："	Keep only evidence-bearing contrasts; otherwise state the academic claim directly
Vague referents/comparatives	"这些东西", "更自然", "更适合"	Name the research object, baseline, and evaluation criterion

AI Density Scoring

Score	Action
>70%	Urgent: immediate rewrite
50-70%	High: rewrite soon
30-50%	Medium: review and revise
<30%	Low: light polish only

Edit Types

1. Delete empty phrases 2. Add specifics 3. Split long sentences 4. Restructure 5. Downgrade certainty 6. Remove redundancy 7. Add missing subjects 8. Replace templates

English-Abstract Tense Check ([Script] LOW)

deai_check.py also flags present-tense reporting verbs (shows, presents, demonstrates, outperforms, …) — but only inside the English abstract. Chinese prose has no tense, so the check is gated to the English-abstract region and stays silent everywhere else. It locates that region across the flagship templates: generic \begin{abstract}, thuthesis \begin{abstract*}, and pkuthss \begin{eabstract}; the Chinese abstract / cabstract environments are skipped, and when several abstracts coexist the English one is chosen by content language. A verb whose subject is a figure/table/equation (Figure 2 shows ...) is exempt. Each hit is a [Script] LOW trace suggesting past tense for methods/results.

See ../writing/tense-guide-zh.md for the judgment-level checklist (which verbs are borderline, why is/are are excluded).

Full details: see ../deai/guide.md