A Trust-Aware Framework for Scoring and Routing in Multimodal Generative AI Systems

Generative AI now sits in the critical path of modern content delivery. A single user request can trigger a summary from a large language model, a synthetic thumbnail, an audio narration, and an automated transcript, each with its own failure modes. In practice, these failures do not remain isolated. A weak retrieval result can surface as an unsupported summary claim, influence a generated headline, and then shape recommendation and search behavior. This paper presents a trust-aware framework that scores each generated artifact before delivery and uses that score to guide routing decisions such as publish, escalate, revise, or block. The proposed framework combines evidence support, source credibility, provenance completeness, disclosure quality, explanation utility, editorial review, calibration, metadata validity, and policy risk into a single composite score. We describe a platform-agnostic architecture, show how the same controls can be applied across text, image, audio, and transcription pathways, and report scenario-based evaluation results on a controlled synthetic workload. Compared with a baseline retrieval-augmented pipeline, the trust-aware configuration reduced unsupported claims, substantially improved provenance coverage and search metadata validity, and increased latency by a manageable amount. We also discuss operational tradeoffs, limitations, and deployment considerations for publishers and other organizations that deliver AI-generated content at scale.