Rules rules rules. Rules rules, rules, rules. Rules. Rules? Rules rules!
I thought the reason we got Bayesian spam filters in the /first/ place was to not have to use rules.
Agreed though, it's getting hella annoying /still/ getting "Confidence is Back" mails after telling Mail that the /last/ hundred mails named that were spam.
Every filtering system has its flaws. Pattern-matching filters can be defeated if you know the patterns. Semantic filters (Apparently Mail.app uses latent semantic analysis rather than bayesian categorisation) can often refuse to learn to recognise obvious spam because of some obscure mathematical similar to real mail.
Which means if you're one of those unlucky folks who receive a metric fuck-tonne of spam, you have to use both. Mail.app supports this by recognising various other spam-filters' custom headers, but it still sucks to have to do it.
Unfortunately Mail.app doesn't apply rules after rendering HTML, so some spammers are breaking up, say, VIAGRA or PAYPAL with <span> and <font> tags into single letters which you can't really guard against with simple rules or word-frequency mapping (Bayesian or otherwise).
Setting up something like Spam Assassin to check SURBS and the other IP/open relay blacklists and then telling Mail.app to trust its headers can make a big difference; I only get a couple spam a week slip through the combination, and they all use the method above.
I know the feeling, except my problem is with this "Julie" character.
Does she know Denise Stuart? I bet they go way back.
Are the messages from "mail delivery subsystem" actually bounces, or do they have a non-blank sender? If they're bounces, then it's well worth setting up a specific filter to drop bounces from forged-sender spam, so that any real bounces will still get through -- in my experience, statistical filtering isn't very good at distinguishing "bounce from forged mail" from "bounce from mail I sent which didn't get through" (probably because the bounces don't quote enough of the text of the mail?).
If Mail.app produces message-IDs in a reasonably unusual format (especially if they include your own hostname or username or something) then it's easy to scan the body of each bounce for anything that could be a message-ID from a message that you've sent. This is usually effective at detecting real bounces. Then drop anything else with a blank return-path on the floor. (I do this by generating message-IDs of a specific form, but that's probably not an option here. Relevant script.)
They are actual bounces from mail that I never sent. Spammers like to flatter me by using my email address occasionally. Yesterday, for example, I got something like 5,000 bounces. That happens every few months.
I have long since accepted that spam has broken email to the extent that I will never again know when a message that I actually sent didn't get through. Fortunately that doesn't happen very often these days.
Maybe I could find a way to tell them apart, but I really just don't care enough.
Fair enough. I get a couple of hundred forged-from-address bounces per day (it seems to be a continual dribble rather than occasional waves), and almost no real bounces, but when I do get one I'd prefer to know about it.
I found that Mail.app's junkmail filter was leaving me wanting something that worked better. I stumbled upon SpamSieve.
From the site:
Spam Sieve gives you back your inbox by bringing powerful Bayesian spam filtering to Mac e-mail clients. It's quick and easy to control Spam Sieve from within your mail client, and you can customize how it interacts with the rest of your message sorting rules. Other spam filters get worse over time as spammers adapt to their rules; Spam Sieve actually gets better over time as it adapts to your mail. By learning from the very messages that you receive, Spam Sieve is able to block nearly all of your spam, without putting your good messages in the spam mailbox.
Seconded. SpamSieve rocks my world. It's a tad RAM-heavy, but it has an astounding success rate once you've trained it enough, and the latest version colour-codes messages according to how spammy they are, so if you need to go through your spam folder looking for false positives, they should all (hopefully) be at one end of the list.
Try JunkMatcher ( http://junkmatcher.sourceforge.net ), a Mail.app add-on with more comprehensive spam fighting features.
N+1, it seems. My Mail.app started filing PayPal mail as junk a while ago; I don't remember how long it took.
Make sure you don't have any PayPal addresses in your address book.
I'm pretty sure you only read text mails? Most of the junk either has empty text or the string 'get an HTML compatible mail program' or junk that filters well using Bayesian.
I use POPFile and pre-trained it on all the junk I'd accumulated that now has a better than 98% rate and picking up spam. Of course it has a tendency to pick up non-spam from people I haven't ever had mail from before or who use word counts and average close to spam. Arguably I don't want email from such folk anyway.
Virtually all the Paypal mail I get is an attempt to get me to participate in a game of Go Phish and its all filtered as junk.
gmail started thinking "Notification of Payment Received" was spam, and that's just not helpful.