Repository navigation
PaperDelta 2.0 makes Agent proposals easier to review against their actual experiment evidence and original manuscript positions. This stable release follows the project's 1.9 → 2.0 version rollover convention. Configuration schema 10, report schema 9 and the existing 22 MCP tools remain compatible.
- Share declared experiment identity, exact source/manuscript hashes, original context, units, selected records and conflicts across direct, staged and batch proposals, CLI and bilingual Studio.
- Distinguish declarations and supplied rationales from deterministic checks. Show sample totals, truncation and paths to the complete evidence.
- Add bounded undo and typed statistical contracts to staged mapping. Preserve input freshness and the last valid draft after invalid actions.
- Preserve binding selections across English/Chinese switching while clearing confirmation. Revalidate compatible 1.9 drafts; proposal imports and stale inputs cannot reuse an earlier confirmation.
- Keep partial startup logs on uvx timeout.
The exact release wheel passed 1,021 Windows tests (three POSIX-only skips), 92 installed-browser cases, 13 isolated starts and recovery from actual public 1.9 drafts/transactions. The source commit also passed all 17 named platform/browser CI jobs. Recorded model actions replay in 60 cases; 13 authored reference/adversarial controls use no model calls.
The separate real Qwen3-8B Q4_K_M study preserves every attempt and first failure. Its first held-out run completed 1/8 mappings and explicitly abstained on 2/4 required cases. These authored tasks do not establish reliable autonomous mapping or independently assessed abstention reasoning. The second development implementation had fewer invalid actions but lower complete-mapping accuracy; both rounds remain published. Proposals still require explicit review and acceptance.
Read the 2.0 guide, study results and exact-commit CI.
pip install --upgrade "paperdelta[docx,pdf,mcp]==2.0.0"The wheel and source archive use MIT. The separate evaluation ZIP retains the original MIT, CC-BY-4.0 and CC-BY-SA-4.0 notices; it does not relicense third-party papers. SHA256SUMS identifies all three archives. Model weights and executable runtimes are not included.
PaperDelta 2.0 让 Agent 建议中的实验声明、实际证据和原始稿件位置可以一起复核。本次为正式发行，遵循项目 1.9 → 2.0 的版本进位约定。配置 schema 10、报告 schema 9 及既有 22 个 MCP 工具保持兼容。
- 直接、分阶段、批量建议、CLI 与双语 Studio 统一展示实验身份声明、来源/稿件精确哈希、原文上下文、单位、所选记录及冲突。
- 区分声明、自述理由与确定性检查；明确样本总量、截断状态和完整证据查看路径。
- 分阶段映射加入有界撤回和统计类型契约；无效动作保留上一次有效草稿，撤回不绕过输入新鲜度校验。
- 中英文切换保留绑定选择并清除确认；恢复兼容 1.9 草稿前重新验证，导入和过期输入不能复用此前确认。
- uvx 启动超时保留部分输出和失败记录。
精确发行 wheel 通过 1,021 项 Windows 测试（三项 POSIX 专项跳过）、92 条安装包浏览器用例、13 项隔离启动，以及真实公开 1.9 草稿/事务恢复。源码提交另通过全部 17 个指定平台/浏览器 CI 作业。60 条模型动作记录重放通过；13 项自建参考/对抗控制没有调用模型。
独立存档的真实 Qwen3-8B Q4_K_M 研究保留所有尝试和首次失败。首次留出运行仅完成 1/8 项映射，在 2/4 项必须拒答案例上明确拒答。这些自建任务不能证明自主映射可靠，也没有独立评估拒答理由的语义质量。第二轮开发实现虽减少无效动作，但完整映射准确率下降；两轮均公开保留。建议仍须明确复核和接受。
pip install --upgrade "paperdelta[docx,pdf,mcp]==2.0.0"wheel 和源码包采用 MIT 许可。单独的评测 ZIP 保留 MIT、CC-BY-4.0、CC-BY-SA-4.0 原有声明，不重新许可第三方论文。SHA256SUMS 标识三份归档；不包含模型权重或运行时可执行文件。