Decision Structure in Repeated LLM Problem Selection Empirical Analysis of 84 Independent Runs

This report presents the third stage of the RUNPORT experimental series investigating structural properties of decision-making in large language models under repeated identical prompts. Seven language models were queried twelve times each using an identical prompt asking them to identify the scientific or technological problem most likely to be solved in the near future. The dataset therefore contains 84 independent responses. From each response three structural elements were extracted: the selected Top-1 problem, the first reasoning factor, and the first three reasoning factors. The analysis reveals two behavioural layers. The selection of specific problems shows considerable dispersion across runs, while reasoning structures exhibit recurring patterns. This suggests that model outputs combine stochastic object selection with comparatively stable reasoning templates.