Challenge the frontier models with your own research.
Idea → ready task in ~1 hour Your baseline becomes the exam
RSI-Anything is the contribution pipeline for OpenRSI Index. It turns your research idea into a ready-to-run harbor task through an automated workflow. The agent asks for your input only when a task-defining choice genuinely needs it.
Clone the pipeline repository and open a fresh Codex or Claude Code session in it, on a frontier model at its maximum reasoning setting. Trust the repository hooks when the client asks: they attest the session, and without them the trajectory cannot be submitted later. Then ask for the proposal-agent skill — one session, one proposal. It walks you through the research question, baseline, evaluation, scope, and compute budget, and drafts the proposal. Review it before you confirm.
Confirm, and the agent posts the proposal as a Task Ideas Discussion for you. Review is automatic and reruns on every edit. When feedback arrives, go back to the same session: the agent reads the Discussion, revises the proposal with you, and updates it in place.
A PASS on the initial check starts the build on its own. The bot only interrupts for a genuine task-defining question, which you answer in the Discussion with /task. At ACCEPTED it goes on to generate a private task repository. Keep the original session open — you will need it once more.
The private repository is a complete workspace: the Harbor task, the read-only RSI-Harness, and the validator and runner skills. First ask the original session to upload the proposal trajectory and Discussion record. Then, on a machine with the GPUs your task declares, validate the environment, run the task through the harness, and upload the completed experiment trajectory.
We review every task for scientific soundness, rigor, and novelty, and independently reproduce it. Approval and trajectory submission do not guarantee inclusion: a task with an unreasonable, unsupported, or irreproducible design may still be excluded.
Leave brief feedback on proposal-agent and the pipeline in this Discussion. With the experiment trajectory pushed, your task is ready for review. Thank you for contributing to OpenRSI Index.
Scope of our RSI tasks
Anything related to a foundation model are in our scope. Vertical foundation model development like medical, legal, 3D vision, or scientific are also welcome.
Important · Most cases GPU access is required to complete a contribution
In most of cases you may need to have the GPU resources required by your task. Choose a task scope, experiment plan, and compute budget that fit those resources.
Contribute a representative, fully open project you helped build. You know its open research directions, and its reward-hacking risks, better than anyone else.
Pick a domain influential open project that is closely aligned with your domain expertise, and design an RSI task around one of its experiments.
Clone the pipeline repository, confirm a proposal with proposal-agent, and it posts the Task Ideas Discussion for you. The agents take it from there.
Start RSI-Anything ↗