How about using the chi-square?

That might be useful but I’m not as familiar with it as I need to be.

Thanks for doing that. It looks like it is easier to shuffle the message and get a lot of ABAA than it is to get ABAB. I can accept that.

See below the symbols, their count, and scores.I highlighted the ones with the big differences between LRTB ABAB and RLBT ABAA. The high count symbols just do not cycle still LRTB ABAB, but several of them do RLBT ABAA including the backward q.

Another idea is encoding both directions. One direction with ABAB for some letters, and one directions with ABAA for other letters. Seems like it would have been difficult to not smudge the paper that way, unless there were multiple drafts. LRTB ABAB 25% explains the cycle scores, but not 26 L17, right?

EDIT:

There was an issue with the implementation which caused it to overvalue stuff like "ABCDE". I am working to replace the graphs in the thread. If any of the interpretations has changed I will let it know.

Everything is updated. The fix mostly compressed the vertical height of the graphs.