I haven't been following cracking development, is anyone aware of a back of the envelope analysis of how much using a preconceived mnemonic reduces entropy versus random letters? I would expect the math has been done, I think all you need are the 1gram counts for the first letter position[1].
EDIT: Based on a quick spreadsheet calculation, I think uniform A-Z each letter is 4.7 bits, and a phrase constructed of random english words each letter is 4.1 bits, so maybe not all that bad. https://docs.google.com/spreadsheet/pub?key=0Ar03cGpoaUJ3dHp...
[1] https://www.google.com/fusiontables/DataSource?docid=1DlRnW1...
EDIT: Based on a quick spreadsheet calculation, I think uniform A-Z each letter is 4.7 bits, and a phrase constructed of random english words each letter is 4.1 bits, so maybe not all that bad. https://docs.google.com/spreadsheet/pub?key=0Ar03cGpoaUJ3dHp...