多模型互查协作流程:设计文档与迭代记录 / Multi-Model Cross-Verification Workflow: Design Documents and Iteration Records
Abstract
本记录公开一套把"多个大模型互相核查"固化为自动流程的程序设计,以及这套设计从一次手工互查实践中形成的过程,包括设计文档、软件规格、全部历史版本、外部审查与专利检索所用的提示词和各轮外部审查的原文。 主文档《设计理由与思路演变》说明:流程的作用不是扩张模型的能力边界,而是把模型已有、但被平台规则、粗心或虚假声称掩盖的能力释放出来;让不同角色回答不同的问题,使核验与构造分开;要声称一个结论可靠,需要模型之外的验证器支撑,做不到的如实声明实际达到的验证层级。文档同时列出与已有研究的对应、尚未检验的假设,以及可替代的实现方式。 公开目的为开源与防御性公开;作者不就其中任何内容申请专利。本项目的设计思想在作者与多个大模型的长期对话中逐步形成,参与的模型为 DeepSeek(手机 App 与网页版所用的模型,以及经 Chatbox 客户端调用的 Pro 模型)、豆包、Claude、ChatGPT;其中数学项目阶段各模型的分工见《前情提要》,外部审查由 ChatGPT 网页版、豆包 Pro、DeepSeek Pro(经 Chatbox 客户端,文中简称 DSP)完成,被采纳的意见已写入文档,审查原文随记录公开。许多早期讨论已无完整记录,各方的具体贡献无法逐项区分。全部取舍与裁定由作者决定;本记录中文档的文字整理、资料检索与原型代码由 Claude(Anthropic)完成。模型均不列为作者。 按本设计编写的程序正式名称为"观者自在 (AlreadyThere)",见 10.5281/zenodo.22969156。 This record publishes the design of a program that turns multi-model cross-checking of large language model outputs into an automated workflow, and how the design emerged from an earlier project in which several models were cross-checked by hand. It includes the design documents, the software specification, all historical versions, the prompts used (for external review and for a patent search), and the original texts of each external review round. The main document, "Design Rationale and Evolution", argues that such a workflow does not extend a model's capability boundary but releases capabilities masked by platform rules, carelessness, or false claims; that assigning different questions to different roles separates verification from construction; and that a conclusion may be called reliable only when supported by a verifier outside the model, otherwise the level of verification actually achieved must be stated. It also lists correspondences with published research, untested assumptions, and alternative implementations. The documents are in Chinese; the main document has an English abstract. The purpose of publication is open sourcing and defensive disclosure; the author does not seek patents on any of its content. The design emerged over extended conversations between the author and several large language models, namely DeepSeek (the model behind its mobile app and web version, and the DeepSeek Pro model accessed through the Chatbox desktop client), Doubao, Claude, and ChatGPT. The division of work among models in the mathematical project is described in the background document; external reviews were performed by ChatGPT (web), Doubao Pro, and DeepSeek Pro (via the Chatbox desktop client, referred to as DSP in the documents), with adopted comments incorporated and the original review texts included. Many early discussions were not fully recorded, so individual contributions cannot be itemized. All design decisions were made by the author; drafting, literature search, and prototype code for the documents in this record were done by Claude (Anthropic). No model is listed as an author. The program built on this design is formally named 观者自在 (AlreadyThere); see 10.5281/zenodo.22969156.