So, having “hold-out” data allows you to have a nonpartisan set of data (so to speak) for you to refer to later, once you have made use of the “train” and “test” sets. That way you can make a fair assessment as to the difference between the new and old (‘hold-out”) sets.

Or if you need to repeat the experiment or model again – then you can refer to the “hold out data”