On Tuesday, July 20, 2004, 4:59:43 PM, Joseph Burford wrote:
Yikes! I'd say 4 to 5 percent FPs is unworkable for a SURBL....
I should have clarified that.
The 4 to 5 percent on close examination is always email that is borderline IMHO. It just didn't end up being tagged as spam in the end. Compared with a .1 to .5 percent FP on the other surbl lists it still isn't that great.
Thanks for the clarification. Even relative comparisons are useful however.
As for being unworkable it depends on your setup. I was tagging DS with a low score, one of the great things with SpamAssassin is that you have things like Bayes to combat FPs and the end result is the ham gets through :)
Ideally we'd like the FP rates to be as low as possible due to 100% hand checking, or some really bulletproof methodologies. Part of the goal is to get the FP rates low enough that people could feel comfortable using SURBLs with an MTA to reject directly at the application transport level. (In other words use SURBLs before SpamAssassin ever sees the message or on a server with no SA.) There is currently some MTA support which people may be using for personal or small company mail, but it would be nice if the data could be robust enough for use at an ISP, etc. We won't get there with data sources that have relatively many FPs.
Jeff C.