| name | rl-standard-launch-iris |
| description | Validate, submit, and observe standard MarinSkyRL training on Iris for dataset-backed rewards without Harbor, Daytona, terminal-bench, or another agent harness. |
Launch standard RL on Iris
Read the selected configuration and current cloud.iris.iris_backend interface. Resolve the runtime profile,
resources, retries, names, capacity, and artifact destinations at execution time.
Workflow
- Start from a clean committed revision. Confirm that the resolved configuration uses a standard
dataset and reward path and contains no agent harness or sandbox environment.
- Validate model, data, topology, batch geometry, reward function, checkpoint policy, artifact
destinations, and registration policy.
- Run the complete launch with
--dry-run. Inspect job identity, runtime commit and profile, resources, trainer
command, input overrides, retry policy, and durable destinations.
- Submit the reviewed command and preserve its resolved configuration with the experiment record.
- Verify worker admission, distributed initialization, data and reward processing, advancing
training metrics, and checkpoints when due.
Safety
- Do not pass agentic-only arguments or secrets to a standard run.
- Prefer launcher-derived defaults unless the experiment requires an override.
- Never patch a live pod or remote checkout.
- Do not cancel a running job without authority; preserve evidence before an authorized stop.
- A running controller state without advancing batches or steps is not proof of health.
