538's model has the (probably reasonable) belief that polling errors don't happen in a vacuum; if there's a polling error in one direction in a state, then other similar states probably have similar errors. Trump winning Florida would imply at least a 3.5 point error based on current polls, so not insubstantial.