A BLOCKCHAIN-INSPIRED MULTI-AGENT TRUST-WEIGHTED CONSENSUS FRAMEWORK FOR REAL-TIME PROPAGANDA DETECTION

THE INCREASING SOPHISTICATION OF DIGITAL PROPAGANDA POSES A SIGNIFICANT THREAT TO INFORMATION INTEGRITY AND PUBLIC TRUST, PARTICULARLY IN THE CONTEXT OF ALGORITHMICALLY AMPLIFIED MISINFORMATION. THIS PAPER INTRODUCES A NOVEL MULTI-AGENT FRAMEWORK FOR PROPAGANDA DETECTION THAT LEVERAGES A TRUST-WEIGHTED CONSENSUS AMONG MULTIPLE LLMS INCLUDING PRE-TRAINED LLAMA 3.2, QWEN2, AND QWEN2.5 EXECUTED IN REAL TIME THROUGH CREWAI ORCHESTRATION. UNLIKE TRADITIONAL SINGLE-MODEL CLASSIFIERS, OUR SYSTEM INTEGRATES AGENT-LEVEL ACCURACY AND DYNAMICALLY UPDATED REPUTATION SCORES TO INFORM WEIGHTED VOTING DECISIONS. A QUORUM-DRIVEN CONSENSUS PROTOCOL DETERMINES THE FINAL CLASSIFICATION OUTCOME, WHILE A RATIONALE AUDITING MODULE EVALUATES THE SEMANTIC COHERENCE AND STRUCTURAL VARIABILITY OF ACTUAL MODEL-GENERATED JUSTIFICATIONS, ENHANCING TRANSPARENCY AND AUDITABILITY. THE MODEL IS EVALUATED ON A LARGE-SCALE DATASET HQP LLM INFERENCES, TESTED ACROSS VARIED INPUT DIFFICULTIES, NOISE CONDITIONS, AND AGENT RELIABILITY PROFILES. KEY EVALUATION METRICS INCLUDE AVERAGE RATIONALE SCORE, DECISION STABILITY UNDER LINGUISTIC PERTURBATION, AND SENSITIVITY TO ADVERSARIAL DISAGREEMENT. RESULTS DEMONSTRATE THAT THE PROPOSED SYSTEM SUBSTANTIALLY IMPROVES CLASSIFICATION ROBUSTNESS, INTERPRETABILITY, AND ADVERSARIAL RESILIENCE OVER BASELINE ENSEMBLE AND SINGLE-AGENT APPROACHES. BY INCORPORATING EXPLAINABLE CONSENSUS MECHANISMS AND ADAPTIVE TRUST MODELING WITHIN A DEPLOYABLE LLM ENSEMBLE, THIS FRAMEWORK ADVANCES THE STATE-OF-THE-ART IN SECURE, SCALABLE, AND INTERPRETABLE PROPAGANDA DETECTION. THE ARCHITECTURE OFFERS A PRACTICAL FOUNDATION FOR REAL-TIME AI CLASSIFICATION AUDITING AND TRUST-GOVERNED.