So, I did a little generator that models comparing alphabetical letters and occasionally gets it wrong. It outputs a CSV file showing the supposedly "greater" letter first. Here's a run with the first 5 letters of the alphabet:
A,B
A,C
D,A
A,E
B,C
D,B
B,E
C,D
E,C
D,E
This has about a 30% error rate.