OCRopus outputs full confidence values in the .fst files that it
produces. Those confidence values are negative log likelihoods.
It does not output confidence values after alignment with the language
model because that's not really a good thing to do; you need the
segmentation alternatives in order to interpret the confidence values
correctly.
Tom