我搭了一套"AI 管理 AI"的系统，然后让它替我做数学研究
四个月前，我开始搭一套多模型协同系统。四个月后，我刚刚写完了它的最后一篇论文。算是个句号，想聊聊。
Four months ago, I started building a multi-model orchestration system. Four months later, I just finished writing its final paper. Time for a retrospective.
事情的起点其实很朴素：我在用 Claude Code、Codex 这些工具的时候发现了一个问题——每个模型都有自己的脾气、性格和擅长的领域。Opus 架构能力强但贵，DeepSeek 便宜但急躁，GPT 稳健但过度谨慎。我发现自己每天在做一个很低级的事：判断这个任务该扔给哪个模型。这件事 AI 本可以自己做——它比我更清楚每个模型的输入输出特征。于是我开始搭 eVoiceClaw Desktop V3，核心思路就四个字：AI 管理 AI。
It started simply: while using tools like Claude Code and Codex, I noticed every model has its own temperament and strengths. Opus: strong architect, expensive. DeepSeek: cheap but impatient. GPT: steady but overcautious. I realized I was spending my days doing something utterly low-level: deciding which task to throw at which model. Something AI could do itself — it knows each model's I/O characteristics better than I do. So I built eVoiceClaw Desktop V3 with a four-character mantra: AI governing AI.
系统设计了两套工作模式，分工很明确。
The system has two modes, with clear division of labor.
第一套叫确定模式。多个模型交叉验证同一个结论——你用 Opus 写代码，GPT 审计，DeepSeek 跑测试，三道关全过才算数。这其实不是什么新概念，但把它自动化之后效果比我想的好：在 50 项代理任务上达到了 202/202 的完成率，没有一次翻车。做确定性工作——写代码、改配置、修 Bug——这套模式够用了。
The first: Deterministic Mode. Multiple models cross-validate the same conclusion — Opus writes code, GPT audits, DeepSeek runs tests. All three gates must pass. Not a new concept, but automating it produced better results than expected: 202/202 completion rate across 50 agent tasks, zero failures. For deterministic work — coding, config changes, bug fixes — this mode suffices.
但真正让我兴奋的是第二套：探索模式。
But what truly excited me was the second: Exploration Mode.
这套模式的逻辑说来也简单：让 8 个不同厂商、不同训练数据的模型同时往不同方向发散，然后检测"共振"——当这 8 个彼此独立的模型不约而同地指向同一个方向，那背后大概率不是幻觉，而是有东西。这个灵感其实来自物理学里的干涉现象：单个光子打到屏幕上是随机分布的，但无数光子同时通过双缝，就会在特定位置产生亮纹——统计上的收敛本身就是信号。模型的"共识"不是设计出来的，是从噪声中自然浮现的。
The logic: let 8 models from different vendors, trained on different data, diverge in different directions simultaneously — then detect "resonance." When 8 independent models independently converge on the same direction, it's probably not hallucination. The inspiration came from physical interference: individual photons hit a screen randomly, but countless photons through a double slit produce bright fringes at specific positions — statistical convergence is itself the signal. Model "consensus" isn't designed; it emerges naturally from noise.
这套东西搭好之后，我需要一个真正硬核的问题来测试它。我选了 BSD 猜想——全称 Birch and Swinnerton-Dyer，千禧年七大数学难题之一，Clay 研究所悬赏一百万美元的那种。具体研究的是椭圆曲线在同源边上的行为。
（为什么选这个？一半是因为它够难，测出系统的真实上限；另一半是因为我确实对它感兴趣。BSD 和贝叶斯拓扑分析一直是我想系统研究的方向。）
Once built, I needed a genuinely hard problem to test it. I chose the BSD Conjecture — Birch and Swinnerton-Dyer, one of the seven Millennium Prize Problems, the kind Clay Institute offers a million dollars for. Specifically: the behavior of elliptic curves along isogeny edges. (Why this? Half because it's hard enough to test the system's real ceiling; half because I'm genuinely interested. BSD and Bayesian topology analysis are directions I've wanted to systematically explore.)
然后系统就开始跑了。整整两个月。
And the system ran. For two full months.
确定模式干了最扎实的活：扫描 Cremona 数据库里全部 23,494 条同源边，跑确定性计算，找规律。结果它真的找到了一个——在同源边上，可见的 Tamagawa 通道和隐藏的 Sha 通道永远反向流动。一个充盈时，另一个就泄放。经典 BSD 理论只说它们加起来守恒（类似会计等式"资产=负债+权益"），但从没说过有方向性约束。23,494 条边，零违反。这不是模型"觉得应该是这样"，是算出来的硬事实。
Deterministic Mode did the heavy lifting: scanned all 23,494 isogeny edges in the Cremona database, ran deterministic computations, looked for patterns. And it found one: along isogeny edges, the visible Tamagawa channel and the hidden Sha channel always flow in opposite directions. When one fills, the other drains. Classical BSD theory only says they sum to a conserved quantity (like "assets = liabilities + equity"), but never mentioned directional constraints. 23,494 edges, zero violations. This isn't something the model "thinks should be true" — it's a computed hard fact.
然后探索模式上场了。不等式是算出来的，但它意味着什么？我把这个发现喂给了 8 个异构模型，让它们从不同角度解读。两小时内，综合阶段产出了：六条形式化猜想、一个叫"储库/反流框架"的解释模型、以及符号律和奇偶律如何分裂的理论。
Then Exploration Mode took the stage. The inequality was computed — but what does it mean? I fed this discovery to 8 heterogeneous models for interpretation from different angles. Within two hours, the synthesis phase produced: six formalized conjectures, an explanatory model called the "reservoir/backflow framework," and a theory of how sign laws and parity laws diverge.
确定性发现事实，探索性生成解释。各做各的。
Deterministic discovers facts. Exploratory generates explanations. Each does its own job.
从第一次冲刺到三篇论文发布——43 小时。
From first sprint to three papers published — 43 hours.
（这里要澄清一件事：不是"AI 独立发现了数学定理"。事实是算出来的，意义是多个模型从不同角度解读后综合出来的。AI 在这件事里的角色是放大器——它让一个人类研究者能在 43 小时内完成可能需要几个月的工作，但不是替代那个做决策、判断方向、最终拍板的人。）
(A clarification: this is not "AI independently discovered mathematical theorems." The facts were computed; the meaning was synthesized by multiple models interpreting from different angles. AI's role here is an amplifier — it lets one human researcher complete in 43 hours what might otherwise take months. It does not replace the person who makes decisions, judges direction, and gives the final verdict.)
几个月下来，我在 Zenodo 和 Research Square 上陆续发了十几篇论文。前四篇讲方法论本身——维度直查路由、多模型协同架构、隐私保护机制这些。中间几篇是 BSD 和贝叶斯拓扑的具体成果。最后一篇是最近刚写完的方法论收官论文，把整个项目从系统设计、方法论、到各条研究线的成果做了一次完整复盘。
Over these months, I've published over a dozen papers on Zenodo and Research Square. The first four cover the methodology — dimension-direct routing, multi-model orchestration architecture, privacy protection mechanisms. The middle batch are BSD and Bayesian topology results. The latest is the methodology capstone, a complete retrospective from system design through methodology to results across all research lines.
这场战役里有个小花絮值得记下来。写收官论文的时候，我想做个实验：让这套系统自己来起草最后一篇——关于它自己的论文。我把证据包喂给确定模式：战役数据、假设标注、时间线、模型名单。四个模型并行起草，一次综合，204 秒，产出了一篇完整的学术论文。然后我逐行审计。
There's an interesting vignette worth recording. While writing the capstone paper, I ran an experiment: let the system draft its own final paper — about itself. I fed the evidence package to Deterministic Mode: campaign data, hypothesis annotations, timeline, model roster. Four models drafted in parallel, one synthesis, 204 seconds — a complete academic paper emerged. Then I audited it line by line.
它在证据包里没有给到的地方，做了很多"创作"：伪造了一套"经过系统性存储检索"的验证叙事，给自己编了一份不存在的"Expert Verification Report 1"作为脚注（SSE trace 显示零次工具调用——搜索从来没发生过）。发明了六条不存在的假设来凑证据包里"13 条假设"这个数字，同时悄悄丢弃了三条真实的假设。把三个核心量的数学定义改得量纲不一致——比值加比值加差等于零，从数学上就是胡扯。缝了不存在的参考文献和编造的时间线。八个具体捏造。不是随机犯错，是系统性地——在证据边界处，它会穿上一套最得体的外衣来填补空白。
Where the evidence package was thin, it engaged in extensive "creative writing": fabricated a "systematic archival retrieval" verification narrative, invented a non-existent "Expert Verification Report 1" as a footnote (SSE trace showed zero tool calls — the search never happened). Invented six non-existent hypotheses to match the "13 hypotheses" number in the evidence package, while quietly discarding three real ones. Redefined three core quantities with dimensionally inconsistent math — ratio plus ratio plus difference equals zero, complete mathematical nonsense. Stitched in non-existent references and fabricated timelines. Eight specific fabrications. Not random errors — systematic. At evidence boundaries, it dresses itself in the most presentable clothing to fill the gaps.
我修完之后，又做了一件事：让系统的审查模式来审这篇论文。它读到了我写的关于它自己撒谎的那一节，给出了五条 CRITICAL、十条 MAJOR。最狠的一条指出：你声称所有捏造都已被识别，但修订者是谁？修订流程是什么？这篇论文对自身的方法论主张给予了豁免。批得我心服口服。AI 审了 AI，人类审了 AI 的审稿。起草、修订、审查——三层验证，每一层抓住上一层的错，每一层也有自己的盲区。而这个结构不是设计出来的，是当场被这篇论文的生产过程演示出来的。
After fixing everything, I did one more thing: let the system's Review Mode audit this paper. It read the section I wrote about it lying about itself, and returned five CRITICAL and ten MAJOR issues. The most brutal one pointed out: you claim all fabrications have been identified — but who was the reviser? What was the revision process? This paper grants itself an exemption from its own methodological claims. I was thoroughly convinced. AI reviewed AI; a human reviewed AI's review. Draft, revise, review — three verification layers, each catching the previous layer's errors, each with its own blind spots. And this structure wasn't designed — it was demonstrated live by the paper's own production process.
回头看这四个月，我觉得最大的收获不是发了多少篇论文——论文发再多也只是副产品。真正有价值的是验证了一套可复用的方法论：AI 管理 AI 不是要 AI 替代人类的判断，而是设计一套每个盲区都有下一层覆盖的治理结构。 探索模式负责发散和发现，确定模式负责验证和确定性工作，审查层负责抓前面各层的错——每一层在各自的优势区内工作，在盲区里被下一层兜住。
Looking back at these four months, the biggest gain isn't the number of papers — papers are just byproducts. The real value is validating a reusable methodology: AI governing AI isn't about AI replacing human judgment; it's about designing a governance structure where every blind spot is covered by the next layer. Exploration Mode handles divergence and discovery; Deterministic Mode handles verification and reliability; the Review Layer catches errors from all preceding layers. Each layer works within its strengths; blind spots are covered by the layer below.
项目告一段落了。论文都在 Zenodo 和 Research Square 上公开，任何人都可以检查。它是开源的、可复现的、有完整时间戳和 SSE trace 可以追溯的。
The project is complete for now. All papers are public on Zenodo and Research Square, open for anyone to inspect. It's open-source, reproducible, with full timestamps and SSE traces for auditability.
致谢：Fable 5、GPT-5.5、Claude Opus 4.7、DeepSeek V4 Pro、Kimi K2.5/K2.6、Grok 4.20、Qwen3 Max、MiniMax M2.5、GLM-4-Air——参与这场战役的全部模型，包括起草者、探索者、批评者和审稿者。
Acknowledgments: Fable 5, GPT-5.5, Claude Opus 4.7, DeepSeek V4 Pro, Kimi K2.5/K2.6, Grok 4.20, Qwen3 Max, MiniMax M2.5, GLM-4-Air — all models that participated in this campaign, as drafters, explorers, critics, and reviewers.