Deepgraph README

The project's README from github.com/billion-token-one-task/Deepgraph. As one markdown file: README.md.

DeepGraph: an autonomous research agent that improves by being judged by reality

The autonomous research agent of JouleBeat, an open-source RSI lab.
It reads the literature, finds open questions, runs the experiments,
and lets reality decide what is true.

Website · Results · Showcase · Architecture · 中文

MIT Python 3.12+ Harness × Benchmark

Why DeepGraph

JouleBeat builds self-improving AI by letting harnesses and benchmarks evolve together.

How an agent works (its harness: tools, memory, workflow, verification) now moves capability by as much as a model generation, and open source has made harness variants cheap to generate. What is still scarce is the benchmark that tells real progress from a lucky score. So we put both in the loop: harnesses compete under benchmarks, and benchmarks compete on how well they predict results that arrive later.

That loop needs a steady stream of real tasks whose answers can be checked. Computational research is exactly that: a conclusion is confirmed or refuted by rerunnable code, held-out data or an independent recomputation, in minutes to days. DeepGraph is where those tasks come from.

How it works

Step What happens
1 Read Harvests arXiv at scale and extracts claims, methods and results
2 Map Merges them into an evidence graph of entities, relations and contradictions
3 Ask Ten structural detectors find open questions in pure SQL, with no LLM in the path
4 Test Turns a question into a pre-registered, budgeted experiment on CPU or GPU
5 Judge An evidence ladder issues a verdict only when the comparison is fair, and the result updates what gets searched next

Every verdict is auditable, and the system audits itself: in August 2026 it found and retracted 11 of its own conclusions that rested on empty model outputs. A search loop is only as good as the filter that decides what is true. Details in Results.

At a glance

Live production data, August 2026.
Papers 24,407 harvested, 7,005 through the full evidence pipeline
Evidence graph 248,441 entities, 735,913 relations, 238 contradiction clusters
Discovery 25,652 mapped research opportunities
Experiments 150 audited verdicts across six research agendas
Compute 921M LLM tokens invested

The JouleBeat stack

Layer What it is Status
DeepGraph Research agent: real tasks in, checkable conclusions out In service, with paying research groups
Harness × Benchmark libraries Versioned harnesses, and benchmarks scored on how well they predicted later outcomes Harness versioning running; benchmark validation in build
Evolution Kernel The engine: propose, evaluate, accept or roll back Open source, pip install evolution-kernel

Quick start

Python 3.12+ and an LLM API key are the minimum; PostgreSQL enables the full control plane.

python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env    # set at least DEEPGRAPH_LLM_API_KEY
export $(grep -v '^#' .env | xargs)
python3.12 main.py      # open http://localhost:8080

Full deployment (PostgreSQL, systemd, remote GPU backends): docs/DEPLOY.md.

Learn more

Contributing

Researchers: bring a computational question at deepgraph.joulebeat.com. Developers: issues and pull requests are welcome; contributions are covered by the CLA.

License

MIT

Deepgraph (billion-token-one-task/Deepgraph)

Searching Database…

1K ·GitHub
Deepgraph
这是什么？DeepGraph 是开源 AI 自我改进（RSI）实验室 JouleBeat 的自主研究智能体。它阅读科学文献、寻找未解问题、运行真实实验，让结果来判定什么才是真的。
这是独立应用，还是更像一条流水线？它是一条端到端的研究流水线：把 arXiv 论文采集进证据知识图谱，用纯 SQL（不靠 LLM）检测未解问题，再把问题转化为有预算约束的实验，由证据阶梯来评判。
怎么把它跑起来？快速上手需要 Python 3.12+ 和一个 LLM API 密钥；若要搭建带 PostgreSQL、systemd 和远程 GPU 后端的完整控制平面，请参照 docs/DEPLOY.md 中的部署指南。
需要先建好数据库吗，还是可选的？最小化快速上手时 PostgreSQL 是可选的，但要解锁完整控制平面、systemd 部署和远程 GPU 后端就必须有它。
它产出的每个结论都可信吗？这套系统靠自我审计而非保证每次运行都正确：2026 年 8 月，它发现并撤回了 11 条基于空模型输出的自身结论。
它究竟面向谁？它为想要可验证计算结论的研究人员而构建，同时也充当 JouleBeat 自身研发智能体框架和基准的迭代闭环。
Evolution Kernel 是什么？它是 JouleBeat 技术栈背后的开源引擎，负责提出、评估、接受或回滚对框架和基准的更改。可通过 pip install evolution-kernel 单独安装。
它在没有 LLM 参与的情况下如何发现未解的研究问题？十个结构检测器用纯 SQL 查询证据图谱来挖掘未解问题，这一步刻意不引入 LLM，以保证确定性和可审计性。
到目前为止 DeepGraph 实际产出了什么？根据其 2026 年 8 月的生产数据，它已采集 24,407 篇论文（7,005 篇完成全流程处理），构建了包含 248,441 个实体和 735,913 条关系的图谱，并在六大研究议程下给出 150 份经审计的实验裁定。
JouleBeat 为什么构建这个，而不直接给模型做基准测试？JouleBeat 需要源源不断的真实、可核验的任务来评判新的智能体框架和基准；计算研究恰好提供了这一点，因为结论可以通过重新运行代码或数据来证实或证伪。
What are similar projects?Raven, awesome-rsi, rrsi, Proteus, wikiskill, AI-Scientist, KISS-Knowledge-Infrastructure-for-Scientific-Simulation, EvoScientist
Which topics is it in?recursive-self-improvement, reproducible-research, self-evolving-agents, ai-research, deep-research, science, arxiv, scientific-computing, llm-agents, papers, knowledge-graph, harness
Creator
billion-token-one-task
Size
85.1 MB
Stars
1K
Updated
CommitOct 16Releasev0.3Jun 11
Age
7mo
Commits · 90 Days