Why should a held-out test set be used only after model selection is finished? (scenario 21)

The correct answer is: To provide an unbiased final estimate of generalization after hyperparameters and architecture choices are frozen. This matches standard ML/AI engineering practice for Domain 1: ML Fundamentals. Distractors confuse adjacent concepts or unsafe operational shortcuts. In production, validate with held-out data, monitoring, and governance appropriate to risk.