I agree that spam is a moving target and that is why anti-spam systems need constant updating. My current system (over the last 30 days) rejected 87% (around 45k emails) and accepted 13%. Of that 13% (6600) around 300 were classified as spam by the bayesian classifier in thunderbird. Around 80 were manually classified as spam and added to thunderbird's rules. The thunderbird classifier probably classified 2 ham messages as spam. I don't know of any ham->spam errors in the initial filtering phase.

Should rspamd be expected to do better, about the same, or worse?