什么是自主科研引擎?
——以及 DeepGraph 与 "AI Scientist" 类项目的差别

自主科研引擎(autonomous research engine)指能自主推进科研全流程的 AI 系统:扫文献选题、设计实验、真实运行实验、做统计检验,直到把结果写成论文格式。它和"AI 写作助手"的本质区别在于交付物:不是一段文字,而是一次真实发生的研究——有实验、有数据、有可检查的证据。

这类系统共同的难题:结论能不能信

从 Sakana AI 的 The AI Scientist、Google 的 AI co-scientist 到 FutureHouse 的科研 agent,"AI 自主做研究"已经从设想变成一个活跃的方向。但所有这类系统都撞上同一个问题:AI 产出的结论,人敢不敢用?这个数哪来的?显著吗?有没有数据泄漏?是不是模型编的?审稿人会问的这些问题,系统自己必须能回答——回答不了,产出再多也只是需要人从头验证一遍的草稿。

DeepGraph 的三个设计选择

DeepGraph 是 JouleBeat 出品的自主科研引擎,把"诚实可验证"当作第一约束来设计:

1. "信"的部分不交给 AI。数字溯源、显著性检验、种子复现、泄漏检查——凡是有确定性标准的环节,由写死的规则代码把关,AI 只负责探索。查无出处的数字直接拦下,不进结论。
2. 负结果照实报,证据不足拒绝成稿。交付物是"已验证的证据包",不是"保证正面成果"。证据撑不起来时,系统会拒绝把它写成论文——这是设计出来的行为,不是缺陷。
3. 服务形态,不碰你的私有材料。你只提交一句话研究方向,不需要上传数据、代码或稿子;证据由引擎用公开数据集自己造,任何人可复查。免费层产出领域证据地图、经查新的 idea、实验方案、CPU 小样信号;付费层产出多种子真实实验、同算力 baseline、可溯源证据包、按会议模板的成稿。

和其他 "AI Scientist" 类项目的差别在哪

一句话概括:多数项目在证明"AI 能自动做研究";DeepGraph 在解决"自动做出来的研究,凭什么信"。两件事互补,我们选了后者作为立身之本。

谁适合用

• 有明确研究方向,但缺算力、工程或统计人手的研究者;
• 想在投入之前先看到领域证据地图和查新结果的团队;
• 需要"审稿人挑不出毛病"的实验方案与证据包的作者。

---

In English: what is an autonomous research engine?

An autonomous research engine is an AI system that carries out the research process itself — scanning literature for gaps, designing experiments, actually running them, applying statistical tests, and writing up the results in paper format. Unlike an AI writing assistant, its deliverable is not text but a piece of research that really happened: experiments, data, and checkable evidence.

Every system in this space — Sakana AI's The AI Scientist, Google's AI co-scientist, FutureHouse's research agents — faces the same question: can anyone trust the output? Where did this number come from? Is it significant? Was there data leakage?

DeepGraph (by JouleBeat) makes verifiability the first design constraint rather than an afterthought:

• The "trust" part is taken out of the AI's hands. Provenance, significance testing, seed reproducibility and leakage checks are enforced by deterministic, hard-coded gates; the AI only explores. A number without a verified source is blocked from the conclusions.
• Negative results are reported as-is, and the system refuses to write a paper when evidence is insufficient — by design.
• Delivered as a service. You submit one sentence describing your research direction — no data, code or drafts. Free tier: a field evidence map, novelty-checked ideas, a rigorous experiment plan, and a CPU pilot signal. Paid tier: multi-seed real experiments, compute-matched baselines, a traceable evidence pack, and a venue-formatted manuscript.

In short: most projects demonstrate that AI can do research autonomously; DeepGraph answers why you should believe what it produced.