If there were a reliable statistical model for how election results were generated in the absence of error or manipulation (a “generative model” for fair election results), one could develop methods that have a known maximum chance of misclassifying a “clean” election as “tainted,” by posing the classification problem as an hypothesis test as described above. The null hypothesis is that the election is clean; in the alternative, it is tainted. One could then construct tests that have probability at most α of concluding that a clean election is tainted, for any desired α between 0 and 100%. If there were a reliable generative model for election results in the presence of error and/or manipulation, one could determine the chance any particular method would misclassify a tainted result as clean and could seek tests that maximise the power to detect fraud or error of various kinds.