Structured Disagreement as an Architecture for Multi-Model AI

Key Points

• This research aims to introduce a structured framework to analyze various AI models' outputs without enforcing consensus or determining truth.
• Developed xCheckAi as a multi-model comparative framework
• Independently queried multiple large language models
• Maintained audit trails to track model outputs and decisions
• Focused on documenting disagreement, agreement, and uncertainty
• The framework allows for transparent decision-making by highlighting diverse model responses
• It preserves auditability by documenting each model’s reasoning process
• The approach supports human oversight in consequential decisions by preventing reliance on a single model's output.