Source Verification as a First-Class Architectural Layer in Multi-Agent AI Deliberation Systems

Citation hallucination—the generation of syntactically plausible but fictitious source references by large language models—represents a particularly damaging failure mode in multi-agent deliberation systems, where false citations propagate as evidence through downstream agents. We describe the Source Verification Service (SVS), an architectural component that addresses this failure mode by operating as an independent background service between the evidence extraction layer and the deliberation ensemble. The SVS performs identifierappropriate resolution across four citation types (URL, DOI, arXiv identifier, ISBN), applies content match verification to detect misattribution to real resources, and automatically applies tiered confidence downgrade rules to evidence nodes in a structured registry based on verification outcomes. Critically, the SVS preserves rather than removes failed citations, maintaining a complete audit trail while marking verification failures at the evidence node level. We describe the SVS design, its integration with the Augle seven-agent deliberation ensemble, and the design principle of preservation over removal that distinguishes this approach from prior citation filtering methods. Companion paper: Kelly, C. & Saxena, S. (2026). “Augle: A Seven-Agent Deliberative Ensemble for Structured Research with Real-World Calibration.” Preprint, May 2026.