spf-discuss
[Top] [All Lists]

RE: Why SOFTFAIL

2004-06-28 05:59:29
[Chris Drake]
Meng: you need to tell people in your spec that "softfail" will cause
all baysian-equipped email clients (inc. all mozilla/thunderbird/etc
ones) to silently eradicate legitimate emails. 

I don't think you understand how bayesian-style filters work. An
"SPF-Softfail" token will be scored statistically, just like any other
token. If the majority of messages with "SPF-soffail" that the user
trains on are spam, then it is true that token will have a very high
spam score. But Bayesian filters take into account as much information
as possible shen making scoring decisions. They do not block based on
one rule, as you suggest. 

Say I have trained on 50 messages with anb "SPF-softfail" token, and pf
these 49 are spam. SPF-softfail will have a spam probability of
something near 0.98. However, for ham messages, and "SPF-softfail" token
won't overwhelm all the "non-spam" evidence in the message. For example,
if I have trained on 30 messsages from Bob as being non-spam, and all of
a sudden he starts sending messages with "SPF-softfail" tokens, his
messages will still probably get through. Why? Because his return
address, and the address of his sending mail server, and his signature,
and a bunch of other tokens are very "hammy" and overwhelm the
spamminess of the "SPF-softfail" token.

Of course, it is also possible that an "SPF-softfail" token will be
neutral or even hammy after Bayesian training, depending on the mix of
email sent to the user.

Maybe you don't, but some other people DO rely upon email in
life-and-death situations, and huge numbers rely upon it for important
financial reasons, so it is up to us to try and restore the concept of
email from it's new "uselessly unreliable" status before it's too
late.  If the email can't go through, and it's important/legitimate,
the sender MUST know.

Dumping something into a spam subfolder folder is not an RFC-violation.
Is it an RFC violation to accept mail that a person simply never reads?
No.


<Prev in Thread] Current Thread [Next in Thread>
  • RE: Why SOFTFAIL, Ryan Malayter <=