There are many variations of this. One I *think* works (but I don't have the skill to check) is that the universe is "sort of" like a simulation, where only macroscopic items have a defined state, but the macroscopic items have defined contents and a defined energy spectrum, and when you arrange to "look closely" at one of those items, it alters the state of the rest of the item in a computationally conservative way, such that you can't detect the difference until you start getting really close to the limits of the simulation, at which point you get results that are statisticly chosen to confirm the conditions of the macroscopic item. So if you split off a bound pair of subatomic particles, they are so pair has defined characteristics, but there is no definition of the components until you look.
Think of it as a way of simplifying the model so that it can run faster on the host computer. The actual "host computer" may not really exist, but if it did, this would be the most efficient way to program it.