You train a linear regression model using OLS. The closed-form solution gives coefficients that perfectly minimize the training loss. A colleague says "since the loss is minimized, the model is optimal." What critical nuance does this claim miss?