His model actually accounts for whether polls were taken before or after an event, and raises and lowers their impact and error margin based on that. Right up to the debate, his model was giving Kamala a <30% chance and it’s only the inclusion of new polls since the debate that have moved her to 50%.
There’s no guarantee you can draw a circle through the bottom of the four legs of a table (opposite legs can be off in the same direction). Also, most floors are not perfectly flat, therefore you can’t assume the floor is at one elevation.