This paper discusses mathematical modeling of the value of particular items of evidence. We demonstrate that such formal modeling has only limited use in explaining the value of legal evidence, much more limited than those investigators who construct and discuss the models assume, and thus that the conclusions they draw about the value of evidence are unwarranted. This is done through a discussion of several recent examples that attempt to quantify evidence relating to carpet fibers, infidelity, DNA random-match evidence, and character evidence used to impeach a witness. This paper makes the following contributions. Most important, it is another demonstration of the complex relationship between algorithmic tools and legal decision making. Furthermore, at a minimum it highlights the need for both analytical and empirical work to accommodate the reference-class problem and the risk of failing to do so.