“Lovelace et al. (2026) introduced an explicit overfitting penalty term to scaling laws for data-constrained training that grows with both the number of data repetitions and the model's capacity ratio (N/U_D).”