A Chinese university lab open-sources a research AI that promises not to fake your data

A Peking University and Yuan Kong AI (元空AI) joint lab has put a self-hosted AI research agent on GitHub, and bound it, at the code level, to one rule: it will not silently invent a result. OpenAI4S, an MIT-licensed open-source project, went up on GitHub on 6 July 2026, with a pure-stdlib core (no third-party dependencies). The most legible thing about it is not the architecture. It is the prohibition.

The agent runs as a persistent Python/R kernel. Write code, run it, keep the variables, plot, return, and the next call still sees the data. That is unusual. Most code-executing agents treat every step as a fresh transcript; the model re-narrates the world each turn and the round-trip is what costs. OpenAI4S treats a research session the way a Jupyter notebook does, which means a literature search, a code experiment, and a final figure can sit in the same process and be edited in plain English afterwards. The kernel is also where the no-fabrication rule lives, in the action layer rather than the prompt.

The rule, as the project states it, forbids four specific substitutions (qbitAI report). The agent may not use np.random to fake an experiment. It may not hardcode a conservation score. It may not pass BLOSUM (a simple amino-acid substitution matrix) off as ESM (a protein language model). And it may not stand a simulated dataset in for a real one. When an external service is unreachable, the protein database UniProt or the structure archive RCSB, the run is required to say so, not paper over the failure. The discipline is more specific than a generic "be honest" instruction: it names the failure modes it is trying to head off.

To see what that buys, follow a protein. The source's walkthrough has the agent pulling a real sequence from UniProt, downloading the matching structure from RCSB, and then sending the folding step off-machine via host-RPC, the project's protocol for dispatching heavy work to a lab's own box. In the demo, that box is an 8×A100 server and the predictor is Protenix, ByteDance's open-source AlphaFold3-class model. AlphaFold3, for context, is DeepMind's 2024 model for joint structure prediction; Protenix is an independent open reimplementation. The result comes back, the agent writes a Markdown report with pLDDT scores attached (the per-residue confidence score, 0 to 100, that AlphaFold-style models output to flag reliable regions), and the figures are versioned so a human can take the notebook over from there. Each of those calls is a real call to a real service, or the run fails honestly. The 8×A100 specification is a demo, not a platform minimum; any GPU box the lab controls will do.

A lab can run an end-to-end AI research workflow on its own hardware, audit every step, and trust that the agent has not substituted a stand-in for a result. The open Skills system, 33 in the project's skills manifest (the project README rounds to "30+"), 14 of which need a GPU, is the lever that makes the thing extensible: a lab can plug in its own data source, its own docking code, its own single-cell pipeline, and the agent will treat it as another tool. The default model runtime is a 9.9-yuan tier of ByteDance's Volcengine Ark (roughly $1.40, approximate, at current rates), and the project is positioned for a Chinese research audience. The commercial parent is Yuan Kong AI's ChatExcel operation.

The source does not yet answer the question a non-beat reader is now entitled to ask. Is the no-fabrication rule enforced in code, or only in documentation? The repository describes it; whether a CI test actually catches, say, an np.random call before merge, the source does not show. The protein walkthrough is one path; the agent also claims to handle single-cell analysis, molecular docking, and literature search, and there is no independent benchmark for any of them yet. "Code-as-Action," the project's term for treating tool calls as code rather than JSON, is also a rebrand risk. Every code-executing agent from OpenAI's own tooling to Anthropic's does some version of this; the difference here is supposed to be the kernel persistence and the hard-coded honesty constraints, both of which are testable claims. As of this writing, no outside lab has published a review.

The thing to watch is the gap between policy and enforcement. If a commit on the repo adds a CI guard that rejects fabricated substitutions, the story gets sharper. If a structural-biology lab outside PKU posts a run on a protein that did not come from a demo, the story gets sharper still. Until then, OpenAI4S is a release with a constructive design choice written into it and a clear place to look next.