One thing that would have been nice to see is a graph with several lines.
X-axis is bits of entropy, Y-axis is recall rate. The individual lines reflect a single combination of (a) formatting scheme (e.g. diceware) and (b) reinforcement mechanism (e.g. type 20x once per day).
I've run this experiment on myself and found that diceware with intra-word punctuation gives the highest entropy at the lowest effort.
English has only got ~1.2 bits of entropy per character. [1]
[1] Table 1 in https://arxiv.org/pdf/1808.04444.pdf
You can measure this with human native speakers, or with a compression algorithm, or with a naive letter n-gram or word n-gram model, and get different statistics -- each of which is correct in the sense of each model's uncertainty about how the text continues.
In the password setting, words can be used in all sorts of ways, some of which are very unpredictable to an attacker, so it's plausible that each word adds at least log₂(len(dictionary)) to the password entropy when it's chosen at random from a dictionary, even when the attacker expects that the password selection method could have used that dictionary.
log(32,909) is about 15. Which is of course 1.5 bits per character, and that's generously assuming that a person will actually randomly pick from all those potential options. So sorry for what I almost commented, but now I have to wonder why the entropy of a word on its own is so close to the entropy of a word in a sentence. Am I missing something? Is it just coincidence?