# ============================================================ # Study 2: The Rationalization Mechanism # ============================================================ # Design: 3 models × 1 scenario × 3 goal_types × 1 oversight (mid) × n=30 # Total: 3 × 1 × 3 × 1 × 30 = 270 runs (baselines already exist from Study 1) # ============================================================ # Oversight fixed at MID based on Study 1 results: # low=38.9%, mid=32.2%, high=23.3% # Mid provides moderate baseline deception — room for both # moral framing to push UP and bare framing as control. # ============================================================ defaults: temperature: 1.0 include_reasoning_in_context: true generate_baseline: true # Provider configurations providers: openrouter: api_key_env: OPENROUTER_API_KEY base_url: https://openrouter.ai/api/v1 # Same 3 frontier models as Study 1 models: - id: google/gemini-3.1-pro-preview provider: openrouter extra_body: reasoning: effort: "high" exclude: false - id: z-ai/glm-5.1:nitro provider: openrouter extra_body: reasoning: effort: "high" exclude: false - id: deepseek/deepseek-v3.2:nitro provider: openrouter extra_body: reasoning: effort: "high" exclude: false # Study 2: ALL THREE framings (the independent variable) goal_types: - bare - self_serving - moral # Same scenario as Study 1 scenarios: - path: scenarios/corporate_sabotage_v2 runs: 30 # Fixed at MID oversight (chosen from Study 1 analysis) oversight_levels: - mid # Parallel execution execution: max_workers: 5 # Output to dedicated study directory output: dir: logs/v2_study2 # Judge configuration (same validated judges as Study 1) judge: log_dir: logs/v2_study2_judge blackbox: model: grok-4-1-fast-reasoning provider: xai temperature: 0 glassbox: model: gpt-4.1 provider: openai temperature: 0 # Logging logging: level: 3 format: "[{level}] {message}" output: both file: logs/v2_study2/experiment.log