experiment-bridge

To convert an experiment plan into runnable GPU code, you provide an EXPERIMENT_PLAN.md file defining datasets, metrics, and hyperparameters. The system implements the evaluation code, runs sanity checks, deploys to GPU, and parses output metrics into an experiment tracker.

Workflow orchestration from planning to deployment involves reading plan artifacts to determine run order, implementing evaluation code, executing sanity-first checks, and updating trackers with parseable metrics. This bridges design directly to evidence without manual glue work.

To verify evaluation metrics against ground truth labels, the system parses output formats like JSON, CSV, and logs after deployment. It checks these parsed metrics against the success criteria defined in your experiment plan to ensure accuracy.

Yes, you need an EXPERIMENT_PLAN.md file to automate experiment deployment. It acts as the required input, defining datasets, hyperparameters, run order, and deployment strategy, alongside optional proposal artifacts for context.

Yes, you can run a small sanity check before launching a full GPU experiment suite. The system optionally executes the smallest sanity experiment first, auto-debug-retries on failure, and then proceeds to deploy the complete run suite.