Compute Validation — verify before computing, learn from cheap runs before committing to expensive ones

The discipline that catches predictable failures before any compute is spent, and turns cheap "smoke" runs into rich diagnostic measurements that predict production behavior. Pairs with compute-strategy (which decides backend) and campaign-orchestration (which manages long-running state).

One-line summary: reason hard before compute, measure smartly during cheap runs, only commit production when both agree.

When to use this skill

You're about to start a compute campaign that:

• Will take hours to days at full scale
• Has a cheaper smoke mode (smaller dataset, fewer steps, shorter timescale, smaller hardware)
• Has potentially predictable failure modes (drift, accumulation, resource exhaustion, parameter sensitivity)
• Costs real money or queue position when it fails late

This is the gate before any HPC submission, ML training run, DFT calculation, large data pipeline, or other expensive compute.