A dataset contains many measurements per machine. The intended deployment predicts failures for machines absent from training. Which validation split best fits?

1. Duplicate each measurement into the other split
2. Keep each machine entirely within one split
3. Use the machine identifier as the known correct target
4. Randomly split individual measurements from every machine across both sets