Needed documentation about Thunderbird's junk filter

Hi,

I am doing a PhD project which requires some development of a variation 
on MNB (Multinomial Naive Bayes) classification. I want to compare what 
I am doing to Thunderbird's junk mail filter, more precisely - the bias 
depending on what part of the document collection the algorithm uses for 
training.

I know it uses a "train on error" policy, but can't find any 
documentation anywhere... For example - is this only for token training, 
or for class prior distributions as well (i.e. - is the spam/ham ratio 
calculated over the whole document collection, or over only the trained 
ones?)

Best,
Pavel