It doesn't map the real world, and it's not supposed to. It's a model, and like all models, it's designed to make sense of the world by reducing its complexity.

But the problem is that it is needlessly complicated, and the complications are what make the decisionmaking difficult. If the question was "You must flip a switch to the left or to the right. If you flip it to the right, one person will die. If you flip it to the left, five people will die. Again, you must flip the switch. Which way do you flip it?" everybody would give the same answer.

Instead, of "you must flip the switch" you get a speeding train. Instead of the switch you get fat people and bridges. Are the five people children and does the fat person have heart trouble? Is the fat person fat because he's greedy, or does he have a glandular problem. Will anyone see me pushing the fat man? And again, what guarantee do I have that a fat man's body will stop a train? How old is the fat man? How young are the kids? Are there suspicious looking people around who, if the train fails to run over the five people, will just back the train up and try again?

I’m not sure you can just assume that five-versus-one is the same dilemma as four-versus-none, since that might not be how people value the lives of others.

I chose my wording very carefully: "you will save the lives of four randomly selected people who would otherwise die." This is the utilitarian bottom line.

But you're right, four versus none is not the same as five versus one. And five versus one in the abstract is not the same as these five versus that one. The details matter. That is the point I was trying to make.