experiment-bridge

To automate GPU deployment from an experiment plan, the tool reads EXPERIMENT_PLAN.md to generate runnable scripts, configure deployment, and collect initial results across GPUs. It tracks execution parameters like MAX_PARALLEL_RUNS and AUTO_DEPLOY.

To turn an experiment plan into executable code, you need refine-logs/EXPERIMENT_PLAN.md and EXPERIMENT_TRACKER.md, plus FINAL_PROPOSAL.md and IDEA_CANDIDATES.md. Configuration flags like BASE_REPO, CODE_REVIEW, and SANITY_FIRST are also interpreted to generate executable experiments.

Yes, you can run sanity checks before full deployment by setting the SANITY_FIRST configuration flag. This ensures initial validation occurs before executing broader automated experiments and collecting results across GPUs.

Parallel execution for automated experiments is controlled by the MAX_PARALLEL_RUNS configuration flag. It manages concurrent GPU deployments and logs progress directly into EXPERIMENT_TRACKER.md during results collection.

Yes, the experiment automation pipeline supports code review through the CODE_REVIEW configuration flag. This allows automated generation of runnable experiment scripts to be validated before GPU deployment and execution.