Safety filter
Also known as: content filter · guardrail filter
A safety filter is a layer that blocks disallowed content before or after the model — a classifier or rule set guarding the I/O.
- Input side: block malicious or off-policy prompts.
- Output side: block harmful or secret responses.
- Usually a small model or heuristic, not the LLM itself.
The role: a safety filter is a tripwire, not a mind. It enforces policy fast and cheaply; the model (and your prompts) do the actual reasoning.