Metadata-Guided Diffusion and LLM-Orchestrated Quality Governance for Time Series Imputation

High-quality time series (TS) data is essential for reliable analytics, forecasting, and knowledge-driven systems. In operational settings, however, TS are frequently degraded by missing values arising from sensor faults, intermittent connectivity, maintenance activities, and extended partial-blackout events. Whilst recent diffusion-based models have improved imputation accuracy, they remain largely signal-centric, make limited use of semantic and operational metadata, and provide little support for data quality considerations. This vision paper introduces a metadata-governed framework for time series imputation that treats data quality as a first-class system concern. The framework combines diffusion-based generative imputation with a metadata knowledge graph (KG), and large language model (LLM) based agents that coordinate imputation workflows and produce auditable quality narratives. Raw time series are retained in specialised storage systems, whilst metadata, relationships, and quality indicators are maintained explicitly in the knowledge graph. This design supports the dynamic extraction of compact, quality-filtered subgraphs that determine which contextual signals may be used to condition the imputation process. We argue that metadata-aware subgraph conditioning, KG-constrained generative imputation, and LLM-mediated explanation would constitute a practical quality layer for time series imputation. Rather than treating imputation as an isolated preprocessing step, the framework treats it as a governance-aware component within KG-LLM pipelines. The paper outlines core design principles and open challenges for building trustworthy and adaptive data systems.