experiment-bridge

Automating GPU experiments requires parsing an EXPERIMENT_PLAN.md file to derive concrete scripts, then deploying them directly to GPU infrastructure for automated execution and result collection.

Code review before execution is managed by respecting the CODE_REVIEW parameter, allowing optional pre-review by GPT-5.4 before the experiment scripts are deployed to the GPU.

Yes, an EXPERIMENT_PLAN.md file is required as the primary input to derive experiment scripts and initiate the automated deployment workflow to the GPU.

The best way to collect initial results is to execute the experiments via the Sonnet model, which automatically captures outputs and saves them to refine-logs trackers for downstream auto-review loops.

Yes, you can use this workflow for academic research pipelines by providing an EXPERIMENT_PLAN.md and FINAL_PROPOSAL.md, allowing the system to automatically derive and run scripts without manual coding.

Results may not save to refine-logs if the result collection parameters are not properly configured or if the initial experiment deployment fails before the auto-review loop begins.