中文学术写作 AIGC 率降低工具 —— MCP Server + 模式检测引擎
Server has 9 tools with Chinese descriptions and reasonable parameter schemas, but critical gaps undermine production readiness. All tools lack English descriptions (LLM-hostile for non-Chinese agents). Parameter descriptions exist but are minimal (avg ~40 chars, well below 72-char baseline). No output schemas documented, LLMs cannot predict return structure. Error handling is absent; tools return generic dicts without recovery guidance. Tool naming is verb-based (mask_, evaluate_, restore_, analyze_, assess_, generate_, build_) but descriptions are technical jargon-heavy, not LLM-optimized. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Schema completeness varies: most tools have input params with types, but no explicit output schema definitions. Composition is reasonable (each tool has one responsibility), but chaining is unclear, e.g., generate_rewrite_plan expects risk_report object but analyze_ai_risk output structure is undocumented.
扫描文本中的 16 种 AI 写作模式,评估 7 项硬约束,输出风险报告。基于 humanizer-zh-academic skill,检测:理论起笔、段末套路、编号逻辑、被动套话、模板化问题陈述、三元并列、冗余总结、模糊归因、填充短语、泛化结论、AI 高频词、回避「是」、过度排比、三步走、破折号密度、加粗滥用。
逐段 AIGC 评分,支持交互式决策。对文本的每个段落进行独立的 AI 风险评估,并提供段落级别的改写建议。
L2 统计层评估(burstiness/TTR/词汇集中度)。计算文本的统计特征,包括词频突发性、词汇丰富度、词汇集中度等指标。
6 维度 60 分制质量评分。基于多个维度对文本质量进行综合评估,包括学术规范性、逻辑连贯性、表达准确性等。
为单段生成 LLM 改写 prompt。基于段落的风险分析,生成一个结构化的 prompt,供 LLM(如 DeepSeek)用于改写该段落。
计算文本的 Type-Token Ratio (TTR) 词汇丰富度,并扫描违禁词。使用 jieba.lcut 分词,清理标点后计算 TTR = 唯一词元/总词元。违禁词库包含:「综上所述」「由此可见」「不可或缺」「具有重要意义」「此案例印证了」。
No output schemas documented. Tools return dicts (e.g., mask_latex returns {masked_text, count, warnings, status}) but LLMs cannot predict structure without explicit schema. Breaks downstream tool chaining and forces LLMs to guess field names.
All descriptions are in Chinese only. Non-Chinese-speaking LLMs (e.g., English-primary agents) cannot understand tool purpose or when to invoke. Descriptions also lack WHEN/WHY context, e.g., 'analyze_ai_risk' describes WHAT (16 patterns, 7 constraints) but not WHEN to call it vs analyze_statistics.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
基于风险报告生成结构化改写计划。根据 analyze_ai_risk 的输出,为每个检测到的 AI 模式生成具体的改写建议。
提取文本中所有 LaTeX 公式/环境并用语义占位符替换。捕获三种模式:行内公式 $...$、区块公式 $$...$$、LaTeX 环境 \begin{...}\end{...}。采用栈式逐字符扫描,正确处理嵌套。
将 mask_latex 产生的占位符无损还原为原始 LaTeX 公式。占位符格式: [INLINE_MATH_xxx] / [BLOCK_MATH_xxx] / [ENV_xxx]。
No error handling or recovery guidance. Tools catch exceptions and return {status: 'error'} with a string message, but do not categorize errors (retryable vs user-fixable vs fatal) or suggest next steps. LLMs cannot self-correct.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are read-only (no state mutation), but this is not declared. Agents cannot infer safety properties and may over-plan or under-optimize.
Parameter descriptions are minimal (avg ~40 chars vs 72-char baseline). E.g., 'text' param in analyze_ai_risk is '中文学术文本' (4 words, no guidance on length, format, or preprocessing). Descriptions lack actionable constraints (min/max length, required preprocessing like mask_latex).