Bayes Evasion Results

Gibberish and chaff easy to pick up with rules: LONGWORDS, HTML_BADTAG, HTML_NONELEMENT

Surprisingly, the real text attack generally doesn't work at present

Removing "invisible text" actually lowers Bayes accuracy

Reason: bad choice of "real text"

    hence there was no calumny too idle ...

(How many people use the word "calumny" in nonspam mail?)