Interactive demonstration
A durable, human-gated auto-research pipeline
Pose a research question. The pipeline reviews the literature with live search, proposes testable ideas and pauses for you to choose one; it then writes and executes an experiment, analyses the results, drafts a manuscript, reviews it, and pauses again for your sign-off. Every stage is checkpointed, and a run may wait at a gate for as long as you like.
How it works
- 01
Literature review
A Google-Search-grounded review with real sources: key papers, methods in use and open gaps.
- 02
Ideation
Human gateRanked candidate studies, each with a falsifiable hypothesis and a small computational test plan.
- 03
Experimentation
The model writes a self-contained Python experiment; it runs in an isolated sandbox, with a repair loop on failure.
- 04
Analysis
An honest results section — verdict, key numbers, caveats. A failed run is reported as failed.
- 05
Paper writing
A workshop-style manuscript that cites only retrieved sources and reports only executed configurations.
- 06
Review
Human gateA strict automated review with 1–5 scores and required changes.
Why the gates matter
In the archived run below, the first manuscript draft reported 10,000 Monte Carlo trials while the executed script ran 200. The reviewer at the gate sent it back; the pipeline now hands the executed script to the writer and reviewer as ground truth, and the revised paper came back correct. The pipeline is designed to surface exactly that kind of discrepancy, not hide it.
Completed runs
Archived runs with every artifact: sources, ideas, code, results, manuscript and review.
Live demonstration
Start a live run
A research question that a small computational experiment could inform. The run pauses at two human gates: you pick the idea, and you sign off on the reviewed manuscript.
Questions in English work best with the current prompts.