Ahh, wow. Thanks. This was totally glossed over in the article copy:

We only use 30-character sequences that correspond to an Euler cycle in the graph shown in Figure 2 (i.e. a cycle where every edge appears exactly once). These sequences have the property that every non- repeating bigram over S (such as ‘sd’, ‘dj’, ’fk’) appears exactly once. In order to anticipate the next item (e.g., to show a performance advantage), it is necessary to learn associations among groups of three or more items. This eliminates learning of letter frequencies or common pairs of letters, which reduces conscious recognition of the embedded repeating sequence [5].