Silent Misalignment in Multimodal Generative System: A Trace-Based Analysis of Cross-Turn State Drift in Text–Image Workflows

This working paper reports a trace-based field analysis of a multimodal generative workflow in which textual coherence and visual artefact integrity diverged across turns. In a long-horizon design scenario (text–image co-generation for a future-city concept), the language layer maintained high semantic continuity and repeatedly confirmed constraints, while the image outputs drifted across iterations—exhibiting palette dominance, region-programme misalignment, and label-only “corrections” without underlying structural compliance. The paper argues that the core issue is not best framed as model “accuracy” or single-turn prompt sensitivity. Instead, it is a governance failure of cross-modal state control: the interaction lacks non-bypassable mechanisms for state-locking, cross-modal verification, and mismatch visibility. The analysis is intentionally vendor- and model-agnostic, treating the observed pattern as a structural risk that can plausibly recur across systems with different epistemic positioning, consistent with parallel false-completion and artefact-mismatch incidents in AI-mediated document workflows. Supporting evidence (full trace logs, screenshots, and time-ordered notes) is preserved. This working paper does not include raw materials; a later revision will include an anonymised trace index and selected excerpts under appropriate disclosure constraints.