entity resolution is hard everywhere. Because the world is dynamic, but the common understanding of "entity" is a static object.

and the only perfect description of the world is the world, just like on a more trivial scale the only perfect description of what a piece of software does is to run it and see what it does.

So the best I know is to find a level of abstraction that captures enough stability to be useful, with enough flexibility to enable the classification to adopt.

In math, phylogenetic trees might be an example; think Dirichlette processes and exchangeable stochastic processes.