if we can, get the unused transition probabilities out of the HMM data structure.
Might have to be merged with milestone 4 (Haskellish types)