[vadimdne]

Thanks for your post. It sent me to the 3 hour rabbit hole figuring out how to maximise bandwidth given those 104 card-tokens (52x2 for face up/down cards) in real life. I wanted the solution to be practical for, say, two people in prison cells. So the math should be really simple.

My best attempt so far is assigning about ~20 tokens to most popular words (the, and, you...), another ~30 to popular trigrams (ing, ion, tio...), another ~30 to digrams (th, er, on...) and the rest to single letters. The number of unique tokens to be adjusted to the occurrence frequency of corresponding words/trigrams/digrams in natural language.

If the encoder runs out of a token, she just skips it (assuming the decoder will calculate this token is already used and guessing to add it) up until the point where the words become unrecognisable by the decoder. Encoder is also free to use synonyms to avoid running of tokens fast.

Not quite strict system, but my rough assessment gives it 80-100 letters per deck. Should be enough to plan a jailbreak :)