English

Review how experiment changes affect an existing research paper. PaperDelta connects declared CSV/TSV/JSON, static Excel and imported experiment evidence to numbers, comparisons and figure provenance in LaTeX/Markdown/Quarto and optional Word/PDF manuscripts, then shows affected locations in an offline report.

Install with Python 3.11 or newer in a virtual environment:

python -m pip install paperdelta
paperdelta demo --out paperdelta-demo --open

The complete original demo ships in the core package. No model key, GPU, TeX, checkout or plotting dependency is required. The changed scenario intentionally reports outdated numbers and a false claim; the demo creates files but does not apply paper edits. Every run uses a new output directory.

Batch binding, explicit review scopes, traceable exclusions and file watching support repeated review. Reports and CLI prompts support English and Chinese. Verified numeric edits have a preview and recovery journal. Seventeen optional MCP tools are available through python -m pip install 'paperdelta[mcp]'.

Quick start · Workflows · Agent guide · Limits

Optional Word review: install paperdelta[docx] and run paperdelta demo --document docx --out word-demo --open. Paragraphs and ordinary tables use native positions; Word files are read-only and unsupported structures remain explicit.

Checks cover declared evidence and supported LaTeX/Word/PDF structures. They do not certify scientific truth or infer a correct mapping from matching numbers alone. The original implementation is MIT licensed; separately licensed paper evaluation sources are excluded from the Python distributions.

简体中文

检查实验结果变化影响了现有论文的哪些位置。PaperDelta 将明确声明的 CSV/TSV/JSON、静态 Excel 和实验导出 证据关联到 LaTeX/Markdown/Quarto 及可选 Word/PDF 稿件中的数字、比较和图表来源，在离线报告中集中展示待复核内容。

使用 Python 3.11 以上版本，在虚拟环境安装：

python -m pip install paperdelta
paperdelta --lang zh-CN demo --out paperdelta-demo --open

核心包自带完整原创演示，无需模型密钥、GPU、TeX、源码仓库或绘图依赖。变化场景 会有意检出旧数字和失效结论；演示创建文件，不自动修改论文，每次需使用新输出目录。

批量绑定、明确审查范围、可追踪排除和文件监听支持持续使用。报告及命令行支持 中英文；验证后的数值修改可预览并通过事务恢复。安装 python -m pip install 'paperdelta[mcp]' 可使用十七个可选 MCP 工具。

快速开始 · 工作流 · Agent 指南 · 限制

可选 Word 检查：安装 paperdelta[docx] 后运行 paperdelta --lang zh-CN demo --document docx --out word-demo --open。 段落和普通表格使用原生位置；Word 文件保持只读，不支持的结构会明确报告。

检查限于已声明证据和受支持的 LaTeX/Word/PDF 结构，不认证科学正确性，也不以数字相同证明 映射正确。原创实现采用 MIT 许可；单独许可的论文评测源码不包含在 Python 发行包中。

PDF and shared manuscripts / PDF 与多稿件

Install paperdelta[pdf] and run paperdelta demo --document pdf --out pdf-demo --open. Original-page highlights, explicit source/export comparisons and shared metrics help find a corrected source whose exported PDF is stale. PDF is read-only; scans and unreliable content remain unverified. No OCR is performed.

安装 paperdelta[pdf] 后运行 paperdelta --lang zh-CN demo --document pdf --out pdf-demo --open。 原页高亮、明确的源稿/导出稿比较和共享指标，可以帮助发现源稿已更新但 PDF 仍过期的 问题。PDF 保持只读；扫描件及不可靠内容不参与验证，不进行 OCR。

PDF guide · PDF 中文指南

Research review Studio / 论文审查工作台

Run paperdelta studio inside a project to inspect evidence, calculate metrics, select LaTeX/Word/PDF positions and explicitly save reviewed bindings. The local browser interface supports English/Chinese, snapshot comparison, ongoing review, declaration maintenance, position repair, durable draft recovery and original-PDF point selection. Batch binding reuses explicit experiment settings across a result table; portable templates and CLI/MCP proposal imports support explicit subset review. No Node.js, cloud account or model key is needed.

在项目中运行 paperdelta --lang zh-CN studio，即可在本地浏览器查看证据、计算指标、 选择 LaTeX/Word/PDF 位置并明确确认绑定。支持中英文、快照比较、日常审查、声明维护、 位置修复、持久草稿恢复和 PDF 原页点选。批量绑定可在结果表中复用明确实验设置， 可迁移模板和 CLI/MCP 提案导入支持明确的子集复核， 无需 Node.js、云账号或模型密钥。

Workbench guide · 工作台指南

Declared statistics

PaperDelta 1.0 checks complete seed sets, mean/SD/SE/n and explicitly declared Student-t intervals. Compound displays and the written confidence level can be bound in LaTeX, Word and PDF. Missing observations remain unknown. Bilingual Studio, templates, terminal guides and read-only MCP share the same contracts. Statistical intervals are numerical approximations under author-declared assumptions; the tool does not infer significance or establish independence.

统计结果需声明完整种子集合、均值/SD/SE/n 及 Student-t 区间方法；缺观测保持未知。 复合显示与置信水平可在 LaTeX、Word、PDF 绑定，中英文 Studio、模板、终端和只读 MCP 共用相同约定，不自动推断显著性或证明独立性。

Native layout in 1.1 / 1.1 原生排版

Linked Word footnotes/endnotes, common merged headers and rotated/cropped PDF pages preserve original positions. Native review remains read-only. The small licensed native study publishes all outcomes, including 4 supported, 4 missed, 0 mislocated and 56 unknown targets in its first 64-target held-out run.

Word 脚注/尾注、常见合并表头与旋转裁切 PDF 保留原位置，原生审查继续只读。 小规模许可原生试验公开全部结果：首次留出 64 项中支持 4、漏检 4、错位 0、未知 56。

Native scope and results · 原生支持与结果

Static sources in 1.2 / 1.2 静态源码

Core-only Markdown and Quarto readers preserve original source positions for literal prose and pipe tables. Run paperdelta demo --document markdown --out md-demo or paperdelta demo --document quarto --out qmd-demo. Shared evidence, statistics, templates, snapshots, reviewed repair and explicit source/PDF exports use the same bilingual workflow. Both formats remain read-only; code, metadata, dynamic includes and unsupported syntax remain unverified. Nothing is executed.

核心包支持 Markdown/Quarto 字面正文与竖线表格，保留原始源码位置。运行上述 demo 即可体验，共用双语证据、统计、模板、快照、经复核修复及明确的源稿/PDF 导出流程。 两种格式均只读；代码、元数据、动态包含及不支持语法保持未验证，不执行代码。

Static-source guide · 静态源码指南