How do you avoid leakage and overfitting in an AI evaluation process?

I record how many tuning iterations have used each dataset and watch for a widening gap between visible regression results, held-out results, and sampled production evaluation. Access controls and audit trails reduce accidental leakage from benchmark answers into prompts or few-shot examples. When a provider model may already know a public benchmark, I avoid making product claims from that benchmark alone. Final rollout decisions combine held-out performance with guarded online evidence so the team cannot optimize indefinitely inside a closed test world.