A new fraud model is deployed to 5% of payment traffic while the existing model serves the other 95%. The team compares error rates and latency before deciding whether to expand the rollout. What deployment strategy is this?