A data science team at a lending company is building a training dataset for a loan default model. Feature values (customer balances, transaction counts) are updated daily and stored in Vertex AI Feature Store. Each training label corresponds to a loan application timestamp that varies per row. During an early prototype, the model performed suspiciously well offline but degraded in production. The team suspects features observed after the application date leaked into training. What should they do to correctly assemble the training data?