Compute Rules

1. Simulations require the authorization described in Experiment Rules. Paid work requires authorization for its scope and resources; do not ask again within existing authorization. Inspect the recipe and resource request before launch.
2. Use the local Mac for development, small tests, plotting and analysis, not CUDA training. Hetzner (ssh hetzner) is the persistent control host: orchestration and monitoring only, not simulations, training or simulation-backed tests.
3. Olorin is shared: inspect availability, select an idle GPU explicitly and keep heavy files under /scratch/em586. On CSD3/Wilkes3 (ssh csd3), run substantial work inside Slurm allocations, never on login nodes. Batch scheduler requests; leave at least 120 seconds between recurring checks.
4. Use the maintained HPC deployment and Modal dispatcher workflows. Keep run/job/call IDs and logs. Investigate ambiguous submissions before retrying; collect and validate outputs before cleanup. Keep credentials out of Git and chat.

Read Compute Access only for the host being operated.