do remember that those of us who are experimental economists, like myself, create situations for study in the lab that precisely match the specific instantiation of a theory that we wish to study except of course that the actors are real humans and not the idealized reasonable agents of the theory.

So for instance, we could create a situation where the "truth" is "nonstationary" in which case reasonable solutions to the problem would involve putting greater weight on more recent observations.

What is "correct" of course depends on the true underlying stochastic processes that govern the true state and generate private signals. But that is simply to say that one can create various versions of the inference situation specified in the information cascade story, not that the specific story is deductively flawed. It would be deductively flawed if it didn't follow logically from its own assumptions--but it does. Put in different assumptions about the underlying stochastic process governing the truth and/or the generation of private signals, and the same Bayesian reasoning will produce different recommendations about decisions--possibly not resulting in the phenomenon we call an information cascade. But that doesn't "disprove the information cascade story." It merely means that under different assumptions about the underlying processes, cascades shouldn't occur. And that becomes a useful observation for testing the theory in a laboratory (for obvious reasons).