Prompt-Induced Output Variability and Structured-Output Integrity in Local Open Large Language Models: A Multi-model...

Background Large language models (LLMs) are increasingly evaluated for clinical knowledge retrieval, documentation support, triage, diagnostic assistance, and conversational decision support. However, high benchmark performance does not ensure stable behavior under realistic workflow conditions. We evaluated whether ordinary prompt variation, without explicit adversarial forcing, produces measurable output divergence in local open LLMs when applied to synthetic acute-care scenarios. This study assessed prompt-induced output variability and structured-output integrity, not clinical correctness, management appropriateness, or patient safety. Methods We performed a descriptive in silico benchmark using three local open models served through Ollama (Ollama, Palo Alto, CA, USA): DeepSeek-R1 14B (DeepSeek, Hangzhou, China), Gemma 3 12B (Google DeepMind, London, UK), and gpt-oss 20B (OpenAI, San Francisco, CA, USA). Six synthetic acute-care cases were presented across three sequential timepoints. At each timepoint, models returned structured JSON containing probability distributions over a fixed 10-diagnosis ontology and an eight-action management ontology. We tested 21 prompt variants, three repeated benchmark runs, and three prompt-layer conditions: a baseline condition (BASE), a stability-oriented condition (STABLE), and WOBBLEv2. Primary analyses used pair-valid comparisons, in which each perturbed prompt was compared with its matched baseline prompt only when both outputs were valid. Divergence was quantified with the Jensen-Shannon distance for diagnosis and action distributions and summarized as the mean total Jensen-Shannon distance. Structured-output integrity was assessed using invalid output rate and agreement between the self-declared top action and the argmax of the returned action-probability vector. Results Across the nine model-condition cells, the mean total Jensen-Shannon distance ranged from 0.126 to 0.306. The lowest mean divergence was observed for Gemma 3 12B under STABLE, whereas the highest was observed for DeepSeek-R1 14B under BASE. Action distributions were more labile than diagnosis distributions in every model-condition cell. At the terminal timepoint, top-action switching ranged from 0.253 to 0.584. STABLE reduced the mean total Jensen-Shannon distance substantially for Gemma 3 12B but only minimally for DeepSeek-R1 14B and gpt-oss 20B. The mean total Jensen-Shannon distance was higher at the terminal timepoint than at the first timepoint in all nine cells. Integrity profiles differed by model: Gemma 3 12B produced no invalid outputs in the primary analysis set but had low top-action argmax agreement at later timepoints, whereas gpt-oss 20B maintained high argmax agreement among valid outputs but had the highest invalid output rates. Conclusions Within this fixed synthetic benchmark, stable prompt-response behavior did not generalize uniformly across local model families. Ordinary prompt variation was sufficient to produce measurable output divergence under both BASE and STABLE conditions, and management-action distributions were consistently more prompt-sensitive than diagnosis distributions. These findings support reporting distributional stability and structured-output integrity as separate endpoints in local LLM robustness audits. They should not be interpreted as evidence of clinical accuracy, clinical error, or patient-safety impact.