A useful scenario for models is for moderation, for biases, violence, hallucinations, etc. Would be useful to add those.