The twelve paired observations are fixed. Line recomputes the least-squares slope and intercept from the selected fitting rows. Nearest neighbor copies the Y of the closest fitting X, with a lower Point ID winning an equal-distance tie. Both are actual fitting/prediction rules, not scripted score changes.

Train only uses four Train rows. Validation scores can guide candidate choice; Test is evaluated later without refitting. All rows (leak) deliberately uses all twelve targets, including the groups named Validation and Test. Their group labels stay visible for comparison, but their evaluation is contaminated. Each group’s MSE still averages its own four errors.

Changing a setting hides Test evaluation and changes the fit when relevant, but cannot make previously observed Test targets new. Repeatedly consulting the same Test results to choose settings gives them a validation role; real final evaluation then needs untouched data. Learned preprocessing also belongs inside the training-only fitting procedure. This toy has no preprocessing operation, cross-validation, automatic model search or confidence interval.

Left training evaluates higher X values outside its fitting range. Appropriate real splits depend on the task, dependencies, groups and time; one fixed partition or low score does not establish representative generalization. Full-precision errors precede three-decimal display. MSE uses squared Y units, not classification accuracy or a causal quantity.