Where? I don't see it anywhere. There's some discussion of total volume perhaps being non-independent from box to box, so an underfilled box is followed by an overfilled one, but this isn't the point the grandparent is making. The grandparent is making the point that there are probably systems in place to prevent a box from being 80% red, say, so that the assumption each individual skittle in a box is independently uniformly drawn from each possible color does not likely accurately model the dynamics.

Author of the article here; you have a point that there is no explicit discussion of validating this assumption, beyond the variability shown in the colored curves in the "count per pack" plot.

Having said that, this small sample is indeed reasonably consistent (or at least not inconsistent) with that iid assumption for the color of each individual Skittle. We would not expect to see any 80+% red packs even assuming that color was perfectly uniformly iid, because the probability of observing such a pack is so small (less than 10^(-19)).

However, still assuming this model, we should expect to see packs with very small proportion of reds... and we do, with one pack having just 3 red Skittles, for example. The entire distribution of proportion of red follows the assumed binomial distribution very closely.