An ML team stores 40 TB of historical training data in Amazon S3. About 15% of the datasets are accessed daily for active model retraining, while the remaining 85% are older datasets accessed unpredictably—sometimes not for months, then suddenly needed for a compliance-driven re-training run that must start immediately. The team wants to minimize storage cost without risking retrieval delays or having to manually manage lifecycle transitions. Which S3 storage class should they choose?