A Deterministic Multi-LLM Consensus Engine for Parent-Facing Interpretation of Pediatric Rehabilitation Therapy Reports: An Anonymized N=1 Feasibility Case Study

This technical case study describes a restricted graph-based engine that was designed to help convert pediatric therapy reports into safer parent-facing home activity guidance. The problem addressed here is practical: therapy reports often contain useful clinical observations, but parents may not always know how to translate those observations into short activities that can be done safely at home.Using a standard large language model for this task is risky. A single prompt may produce a fluent summary, but it may also invent goals, miss important triggers, or generate advice that is not clearly linked to the original report. For that reason, the proposed engine does not rely on one open-ended LLM response. Instead, it converts the source text into an auditable Directed Acyclic Graph (DAG), separates the work across role-limited processing agents, and requires each agent output to follow a defined JSON structure.The main purpose of the engine is to make every parent-facing recommendation traceable. Each recommendation must be linked back to source evidence, a therapy objective, safety checks, and a review decision. The case study is based on two de-identified pediatric therapy progress reports linked to one anonymized profile (N = 1). The study shows how the engine can reject out-of-scope recommendations, send uncertain cases to human review, and produce low-burden home activities that remain connected to the original therapy documentation. This is a technical feasibility case study only; it does not claim clinical efficacy or therapeutic validation. Keywords—Large Language Models; multi-agent systems; pediatric rehabilitation therapy reports; Applied Behaviour Analysis; Occupational Therapy; clinical NLP; AI safety; human-in-the-loop; evidence grounding; healthcare informatics.