Most moderation bots still start from a list of words. It's simple, fast and predictable, and it fails in two directions at once.
Too strict: it can't see the room
"you're so bad at this"
Between two friends mid-game, that's banter. Aimed at a new member who just asked a question, it isn't. The words are identical, so a word list has to pick one answer for both. Either it flags your regulars all evening, or it lets real pile-ons through.
Sift reads each message with the few before it. Same words, different conversation, different call.
Too loose: scammers read the list too
Spam and scam accounts have spent years learning to dodge filters. Letters get spaced out, swapped for lookalike characters, or split across three messages. Links point to domains one letter off from the real thing.
Sift joins split messages back together and maps lookalike characters back before checking anything. Links are compared against the domains they're imitating, so dlscord-gift.com is caught the moment it's posted, with no smart check needed.
Mods still make the call
Context helps, but it isn't perfect. So Sift acts on its own only when it's confident. Everything in between goes to your log channel as a case card with the message, the reason and the member's history, and one button to handle it. The overturn rate in /stats tells you how often your team disagreed.