MesQuilla

There are lots of oddities in the tokens if you really take the time to examine them. I can't recall a case though where I examined the "why" of a particular token and decided that there was a true bug in the software. Most are training issues.

What that token is telling you is that, of the messages that you have trained, a much higher proportion of those that were sent to: were junk than good. Even if "a large proportion of the emails that I get (and want)" are like that, an even LARGER portion of those that you get (and don't want) are like that.

Now that could be real, or it could be an issue of inadequate training. For training, you should have about the same number of junk as good messages trained. If it is real though, then the junk analyzer is biased toward marking as "junk", and it should be. Tinkering with that one token is not likely to improve the overall performance of the filter. Look at it this way: a message with no real content, but addressed to you, has a 94% chance of being junk. Think something like image spam here.

On my own system, I am now experimentally using a filter (from FiltaQuilla) to automatically train messages as good that match my address book. I need though to add the capability to limit the number of trained good to be about the same as the number of trained junk, as there can also be issues that arise from too much training of good (particularly if the messages differ in some systematic way from the junk, such as being at different times.)