Sync them right next to each other, then move one of them. The other way you could test this theory is to have one clock tell the other the time over an optical link and then have the other do the same. If the speed of light was different in different directions. Each would measure a different lag.
Well, moving them is out of the question, since, you know, motion will change the clocks time. If you re-sync them, you bake the “error” into your framework. If you try a timer, the timer is offset. If you try and propagate a signal, the signal is offset. And eventually, you have to compare the two times, which muddies the waters by introducing a third clock.
Basically, there is no way to sync two clocks without checking both clocks, ergo, no way of proving or disproving. That’s the premise.
In practicality, I assume it is constant, but it’s like N=NP. You can’t prove it within the framework, even if you really, really want to believe one thing.
If you move one clock very slowly away from the other, the error is minimised, perhaps even to a degree that allows for statistically significant measurements.
To cite the Wikipedia entry that one of the other commenters linked:
“The clocks can remain synchronized to an arbitrary accuracy by moving them sufficiently slowly. If it is taken that, if moved slowly, the clocks remain synchronized at all times, even when separated, this method can be used to synchronize two spatially separated clocks.”