Antoine Nasser and Alistair Adcroft investigate how machine learning can be used to develop stable and accurate numerical schemes for solving the linear advection equation. Their study identifies the key factors governing the performance of data-driven finite-volume methods, including network architecture, training data, and normalization strategies. They show that data-driven reconstructions based on cell averages are shape-specific, limiting their ability to generalize across different classes of solutions. They introduce a machine-learned flux limiter that improves shape preservation relative to widely used classical schemes and demonstrate that training on polynomial profiles yields stable, high-order accurate discretizations. Overall, the work provides practical guidelines for designing robust and generalizable machine-learning-based numerical methods for scientific computing.