1, * Faijan KHAN, 1 Chad PEIPER, 1 Amir JABERZADEH, 1 Afaan SHAIKH
2 and Jason GENG
1 Bayes Solutions, 840 Apollo St, El Segundo, CA 90245, USA
2 International Data Engineering and Science Association (IDEAS), Los Angeles, CA 90013, USA
* E-mail: faijian@bayes.global, peiper@gmail.com, amir@bayes.global, afaan@bayes.global, jason@joinideas.org
Received: 2 Feb. 2026 /Revised:27 Feb. 2026 /Accepted:2 Mar. 2026 /Published:23 March 2026
______________________________________________________________________________
Abstract: This article presents Autonomous Trust-Weighted Consensus for continuous learning in decentralized Retrieval-Augmented Generation systems. The proposed framework extends prior decentralized retrieval architectures by introducing a formal trust update mechanism and weighted validation rule that dynamically adjusts contributor influence over time. The system integrates content-addressed storage, distributed vector indexing, and validator sampling to enable transparent contribution management while mitigating adversarial manipulation. A theoretical robustness analysis establishes sufficient conditions under which the influence of malicious participants decreases over repeated interaction cycles. Experimental evaluation is conducted across two domains: immigration policy documents and scientific research abstracts. Results show that the proposed approach preserves retrieval accuracy under clean conditions while significantly reducing performance degradation under controlled data poisoning attacks when compared with majority voting and static reputation baselines. Additional experiments evaluate recovery time, scalability, and statistical significance across multiple seeds. Reproducibility artifacts, configuration details, and experiment scripts are provided to support independent verification.
Keywords: Decentralized retrieval-augmented generation, Trust-weighted consensus, Continuous learning, Adversarial robustness, Distributed validation, Vector database governance.
______________________________________________________________________________
