experiment-bridge

To automate turning an experiment plan into runnable GPU code, you provide an EXPERIMENT_PLAN.md file to trigger automatic code generation, GPU deployment, and result collection. This workflow closes the loop between experiment ideation and evaluation seamlessly.

Before automating experiment deployment, you need to prepare an EXPERIMENT_PLAN.md file outlining your experiment design, and optionally a FINAL_PROPOSAL.md. These files provide the necessary context for the system to generate runnable code and deploy it to GPUs.

Yes, you can run parallel GPU experiments and aggregate the results automatically. The workflow supports parallel runs and result aggregation, feeding the collected data into auto-review workflows and updating refine-logs trackers with the initial outcomes.

Automated experiment deployment integrates with existing research logs by updating refine-logs trackers with the collected results. It also integrates with the /run-experiment command, ensuring that deployment outcomes are systematically recorded for subsequent review cycles.

A key limitation of automating experiment implementation is that it requires an existing EXPERIMENT_PLAN.md to function, meaning it cannot generate experiments from scratch. It is designed to deploy and collect initial results rather than handle full-scale, long-term experiment management autonomously.