A DevOps engineer manages a payment processing service on EC2. The team wants a CloudWatch alarm on the p99 latency metric (published every minute) to notify the on-call engineer via SNS when the service is genuinely degraded. During load testing they observed that a single elevated 1-minute datapoint occurs frequently due to garbage collection pauses, but these are not real incidents. Real degradation is characterized by latency staying high across most of a 5-minute window. Which alarm configuration best reduces false alarms while still catching sustained degradation quickly?