When evaluation conditions differ significantly from real-world deployment, making benchmark results less predictive.