I'd illustrate the issue with some sort of genetic landscape plot, where all the little details such as "fitness" and likelihood to survive during current conditions is based on a height in the landscape. Now, imagine that you have peaks in this landscape, where breeding and survival rates are higher than everywhere else. Then it follows that the closer individuals of a species is to the peaks in their particular fitness landscape, the more likely they are to pass their genes onwards. This part of the theory is fairly simple; but the problem is the landscape. It will be constantly changing due to billions of factors entering into survival rates/ general fitness - some factors can be rather large, like availability of nutrients and presence of predators (in polar bear form or viruses, it doesn't matter much) and certain other environmental conditions, but a lot of the things are factors you can't really consider unless you are omniscient (red feathers, hard beak )... So, in short, it is extremely complex to find the fitness landscape for a species at any given time, and as such, it is all but impossible to design an experiment or observation that could falsify the initial assumption. Suffice to say that there has not yet been any conclusive evidence to say that it doesn't work...

Now, to take a Popperesque view of science in this case might be somewhat flawed... complex biological systems simply have too many factors to consider; Poppers views are more properly applied to science where you can churn the numbers for everything involved instead of doing large scale simulations based on approximations.