Choose an operator task

These task-oriented guides cover the work that surrounds simulation and network training: discovering resources, submitting jobs, preserving lineage, checking candidate artifacts, and promoting a network safely.

Use the quick-test configs while learning. The files under configs/examples/ are templates for larger runs, not universal production settings.

Prepare and submit work

• Configure your cluster resources to discover personal lanes without committing personal allocations.
• Generate and submit Slurm jobs to rehearse scripts, select resources, fan out generation, and preserve submission results.
• Track runs with MLflow to keep generation, training, and promotion linked through one authoritative store.

Promote an artifact

• Validate and inspect a candidate network with the four automated gates and the optional visual inspector.
• Run a parameter-recovery sweep to test whether a network supports inference, not just whether it is a well-formed density.
• Stage and publish a validated network to isolate one run, inspect a dry-run plan, and upload only to a staging repository.

The promotion guides deliberately stop short of the production Hugging Face repository. lan-publish refuses that destination unless given --allow-production and an interactive confirmation, so production promotion remains an explicit, separately governed action.