One alternative to conventional DBEF involves “learning” the distribution of irregularity-free results from the data. This is especially appealing if there is a set of elections available that is believed to be accurate, for elections that took place under conditions similar to those of the election under scrutiny. This allows the strong distributional assumption of conventional DBEF to be relaxed: the irregularity-free numeral distribution does not need to be known ex ante, and does not need to be Benford’s Law. Instead, the test relies on weaker assumption that the numeral distributions for accurate elections differ from those of elections with irregularities.