Interesting, but I am not convinced that the accident data can _really_ correct for things like demographics of drivers. Moreover, there does not seem to be anything in the way of error measurement. In other words, how do you know that a score of 101 is actually worse than a score of 99?
I suspect that accident data doesn't compensate for driver demographics. This is imprecise because of a lack of data, so scores that are 'close' are likely within the margin of error.
To use your example, a score of 101 would be close enough to 99 to be indistinguishable. However, a score of 140 would be distinct enough from 99 be noticeably worse or better.