exp-pilot-run

A pilot experiment using a spec YAML generates code and requires human confirmation before remote execution begins. This human-in-the-loop safety gate ensures safe diagnostic training runs by letting you inspect configurations prior to local GPU or remote SSH deployment.

To collect raw training evaluation metrics, you execute a pilot run for a specific idea slug. The process outputs raw JSON results and a pilot log under a dedicated code directory for baseline comparison and diagnostic reporting.

Running a pilot experiment does not modify the wiki or make automated pass/fail judgments. It strictly focuses on generating deterministic pilot artifacts and collecting raw run data while avoiding automated conclusions or graph edge creation.

Yes, research automation for pilot experiments supports both local GPU and remote SSH deployment scenarios. The system orchestrates these environments while pausing for user confirmation before executing the generated pilot code.

The best way to ensure safe diagnostic training runs is using a human-in-the-loop safety gate requiring manual inspection of generated code. This approach prevents unintended side effects by avoiding wiki modifications and automated pass/fail judging.