ghost-annotator

The Ghost Annotator framework combines conformal prediction with collaborative filtering to model LLM behavior and human label variation in content moderation, revealing structural demographic biases in larger models.