Niklas Hambüchen
Hi Ben,
Could we not have a captcha instead of a reject, to avoid false positives? That would require no training.
Since I assume most Trac spammers are extremely unsophisticated, a simple hardcoded question like "What programming language is GC all about?" may be sufficient.
The CAPTCHAs being broken are the reason why this incident occurred. I have added some more CAPTCHAs to try to dilute the pool of answers that they already know, but they still seem to solve them easily enough regardless. I can only imagine they have some sentient beings sitting at computers solving CAPTCHAs. I don't really feel like we can make the CAPTCHAs themselves any more difficult without excluding real new users, which I really want to avoid. Regardless, my goal here is to error on the side of less filtering, not more, even if this does mean more manual maintenance. To this end, I've configured the filters such that the probability of legitimate activity being suppressed should be negligible, * I've been careful to only train the Bayes filter on obvious spam; I have tested it against various snippets from the wiki and mailing list and have yet to see it score anything legitimate with a spam likelihood > 5%. * Even if the Bayes filter does deem your content to be spammy enough to warrant further attention, you will merely be asked to solve a CAPTCHA. Posts will not be outright rejected unless it is quite clear that they are spam. I am optimistic that the filtering will have negligible effect on legitimate traffic. As a smoke test I managed to create a new account, open a new ticket, and start a new Wiki page without even needing to solve a CAPTCHA. Cheers, - Ben