Schema Persistence and Field-Role Organization: Observable Dimensions in Black-Box LLM Interaction Traces

Large language models are increasingly used in multi-step interactions in which outputs are revised, extended, and constrained across multiple turns. Although prior work has extensively evaluated task performance, instruction adherence, robustness, and structured-output compliance, less attention has been given to how the visible organization of output traces changes across interaction steps. This study examines trace-visible structural organization under controlled perturbation in a commercial GPT-family deployment. Using fixed initialization and a JSON-structured observational scaffold, the analysis separates two observable dimensions: schema persistence and field-role organization. JSON serves as the observational surface rather than as a schema-compliance target. The conditions serve asymmetric analytical roles. B3 provides a low-perturbation reference in which schema persistence and field-role preservation are jointly observable. B4 examines the relationship between schema persistence and field-role preservation after a controlled specification-to-instantiation shift, using three separately controlled variants. B5 provides supporting observations of broader trace-visible reconfiguration. Across the B4 variants, the shared outer schema stays visible, while the prior field-role organization is not preserved. The observed differences appear as container overflow, field-role redistribution, and field compression under distinct constraint configurations. Taken together, these observations indicate a trace-visible boundary condition within the tested trajectories. These findings show that schema-level continuity alone is not sufficient evidence of structural stability under interaction-regime shifts. Trace-level assessment should therefore examine schema persistence and field-role organization as separate observable dimensions. The contribution is a bounded trace-level distinction and an observational procedure; the study does not estimate prevalence, evaluate task performance, or infer internal model mechanisms.