One of the differences about Silver's model is that it generates both a most likely result and a 90% confidence spread. Both of those are important to the overall simulation, and to the individual race chances.
For example, for SD he shows a central value of R+12 and for IL the central value is D+9. Naively one might expect the Pub in IL to have a significantly better chance than the Dem in SD. However, the 90% confidence limits in SD are plus or minus 16, while in Il they spread is plus or minus 13. The result is that there is about the same spread that crosses the center line, so the probability of an upset is comparable, 13% in SD and 11% in IL. The difference between those requires knowledge of the values to at least one decimal place which isn't given on the graphic.
Speaking of difference in upset chances, I would note that there is little statistical difference between 11% and 13% given the size of the 90% confidence spreads in their model. Similarly there is little difference in 52% or 54%, or even if they were flipped the other way. Based on the spreads, I would round off the chances to the nearest 10%. However, small changes in polling could cause apparently large jumps, such as a shift from 52% to 56% would be magnified to a shift from 50% to 60%. Given the wide and generally less sophisticated readership on matters of statistics, I would expect that a 10% jump would get far more attention than it would deserve. That might be part of the reason that they are reporting the accuracy of the chances in each race to such a degree - it diminishes the reactions to normal statistical shifts in the data.
