It seems that you are creating links with equal weight, which is not how Google’s algorithm works. If you study the original Backrub document you will discover that PageRank is a weighting factor. Therefore, a small number of links may often pass a great deal more influence on relevancy scores than a larger number. Then there are Trust factors that also appear to weigh the value of anchor text.
I don’t see in your article where you discuss the methodology you used to isolate these factors and normalize the data. If you have not normalized the data you may have drawn an incorrect conclusion about anchor text value. I see no reason to believe there is any change in anchor text value based on your analysis as presented.
Signals of relevance cannot be separated from their weighting factors to determine their overall value in the ranking. Web pages are ranked based on a formula that combines multiple signals of relevance weighted by factors such as PR and Trust. When you try to look at raw data from an individual signal and exclude the weighting influence of other factors it can lead to all sorts of erroneous conclusions.
I believe that if you were to repeat your analysis on a number of smaller subsets of data you will see wildly varying results. If you cannot consistently repeat the experiment and achieve the same results how can you claim your conclusion is valid?
