AISmithLab/CoBRA

Implements a closed-loop system with a Cognitive Bias Index that measures agent behavior against classic social science experiments (Milgram, Asch, Wason) and a Behavioral Regulation Engine offering three control methods—prompt engineering, representation engineering, and fine-tuning—to achieve quantifiable, reproducible bias levels across models. Supports four validated cognitive biases (Authority, Bandwagon, Confirmation, Framing) with standardized 0-4 control scales, enabling researchers to operationalize social science knowledge as reusable experimental environments for LLM agents.