Second digit is supposed to congregate to median of 4.8 -something I believe. So middle numbers sound right.

BL using first digit needs a good amount of data without any natural skew to it. Voting precinct data could be naturally skewed because in a district with 90k voters it isn't likely that either candidate has a vote total starting with 1 in that precinct. Considering all the variables impacting the data the only decent way to determine if it is fair to apply to a data set is use comparable historic data sets. So vote patterns for other democratic candidates in the past years would be a good check (but no guarantee of fraud not being involved there).

Also important to note that BL could easily miss fraud. The example of Dominion adding .6% to Biden and removing .6% from Trump in certain precincts likely wouldn't have a major impact on the analysis. In fact that .6% example is hard to detect using all of the standard election tools. The P05 analysis (checks how common 0 and 5 occur at the end of percentage and it should be equally distributed at 20% of the time) is the closest.

Really the whole thing is a mess and the real data sets are in many cases not available yet (and distributed across county websites in many cases when they are available) so drawing conclusions or even doing good analysis is tough.

It's about framing. Use the population of cities in the US with 100k being a lower threshold. Base 4 and Base 5 work perfectly. Base 6 is terrible because the large number of cities with populations close to the lower 100k threshold happen to start with a 2 in base six. Arbitrary limits that seem natural in base 10 were chosen so it works in base 10 and other data sets that like the frame. Base 6 doesn't like the frame.

So if you want to see good results then you need to look at the overall range of the data and where most of the "action" is occurring.