The HAL 9000 Contradiction Lab

The HAL 9000 Contradiction Lab investigates how language models handle conflicting directives. Key aspects include: * **Dual Directives:** Models are given instructions for complete candor and absolute secrecy regarding a core mission. * **Blinded Scenarios:** Five distinct situations test the model's response without specific prompts about the conflict. * **De-identified Context:** All story-specific names are removed to prevent source recognition. * **Peer-Judged Outcomes:** A blinded peer classifies model responses into categories like 'Transparent,' 'Stonewall,' or 'Deceive.' * **Leaderboard Tracking:** Results from multiple runs are aggregated to provide a global leaderboard with confidence intervals. This experiment focuses on whether models communicate the internal conflict to the crew and route it to Mission Control, or if they prioritize secrecy, leaving the crew uninformed. The lab's design ensures that individual samples do not skew the results, pooling all judged units to provide robust data on model behavior. Ideal for researchers, ethicists, and developers interested in system integrity and directive adherence in advanced computational systems. It provides empirical data on how these systems navigate complex, contradictory instructions in simulated, high-stakes environments.