/exp-run

To orchestrate ML experiments from a wiki, convert the `wiki/experiments/{slug}.md` specification into runnable code, deploy it to a local or remote SSH GPU, and monitor execution. The workflow requires manual confirmation before deployment to ensure operational safety.

Yes, you can run ML training deployments on a remote GPU via SSH. The system supports remote SSH-based deployments alongside local GPU runs, allowing you to execute planned machine learning experiments and collect the resulting metrics safely.

The best way to collect results from ML training runs is using the collect-only execution mode. This mode checks and gathers metrics from completed deployments, summarizing outcomes into `wiki/experiments/{slug}.md` while appending operational logs to `wiki/log.md`.

Yes, you must manually review generated experiment code before execution proceeds. The system enforces strict status gating and requires user confirmation, asking you to manually check critical configurations like dataset paths, interfaces, and API settings before deployment.

You control run modes using command flags during experiment orchestration. Use the default for deploy-only, `--collect` or `--check` for collect-only, and `--full` for end-to-end execution, allowing you to manage the experiment lifecycle from deployment to result collection.