A small built-in slur list catches common hate slurs and obvious variants on the spot. This is separate from the smart check's hate-speech read, which handles context and less obvious phrasing.
What it is for
Zero-latency removal of slurs that should never stay up while a model request runs. Edge evasions (lookalike scripts, zero-width characters, underscores) may still reach the smart check instead.
What it is not
It is not your whole hate-speech policy. Borderline harassment or coded language still goes through the smart check and your thresholds.