It sounds like this would provide some testable predictions, am I understanding that correctly?
Yes. We would be able to set up experiments where one interpretation would assign non-zero chance to an outcome that has zero quantum amplitude. Run the experiment and observe the outcome enough times and if the positive-chance-but-zero-amplitude event ever happens, then quantum states are merely a calculational tool. If the zero-amplitude event never happens, then as the number of repetitions goes to infinity, you can be as close to certain that the quantum state is real as you'd like. It's worth mentioning that almost all physicists would believe the lat...
From a recent paper that is getting non-trivial attention...
From my understanding, the result works by showing how, if a quantum state is determined only statistically by some true physical state of the universe, then it is possible for us to construct clever quantum measurements that put statistical probability on outcomes for which there is literally zero quantum amplitude, which is a contradiction of Born's rule. The assumptions required are very mild, and if this is confirmed in experiment it would give a lot of justification for a phyicalist / realist interpretation of the Many Worlds point of view.
More from the paper:
On a related note, in one of David Deutsch's original arguments for why Many Worlds was straightforwardly obvious from quantum theory, he mentions Shor's quantum factoring algorithm. Essentially he asks any opponent of Many Worlds to give a real account, not just a parochial calculational account, of why the algorithm works when it is using exponentially more resources than could possibly be classically available to it. The way he put it was: "where was the number factored?"
I was never convinced that regular quantum computation could really be used to convince someone of Many Worlds who did not already believe it, except possibly for bounded-error quantum computation where one must accept the fact that there are different worlds to find one's self in after the computation, namely some of the worlds where the computation had an error due to the algorithm itself (or else one must explain the measurement problem in some different way as per usual). But I think that in light of the paper mentioned above, Deutsch's "where was the number factored" argument may deserve more credence.
Added: Scott Aaronson discusses the paper here (the comments are also interesting).