Simulating Patient Resistance in silico: Evaluating Concepts in Rare Diseases using Patient Agents with Limited Rationality

The evaluation of new medical concepts and digital health devices for rare diseases is hampered by the scarcity and vulnerability of real-world patient cohorts.Synthetic sampling (Silicon Sampling) using personas driven by large language models (LLMs) offers a promising digital evaluation framework, but one that is biased by 'algorithmic perfection': over-rational decision-making, omniscience via retrieval-from-sources (RAG) and sycophancy.This article establishes a methodological framework based on the meta-ergonomics of falsification and the Adversarial Cognitive Twin (ACT) architecture.Rather than increasing the persona's theoretical intelligence, we propose to falsify the synthetic agent by injecting it with calibrated cognitive impairments, semantic entropy, reduced user knowledge bases and multiagent contradiction loops.This methodological constraint enables the simulation of mental load, doubts and user resistance -characteristics specific to clinical reality.An experimental benchmarking protocol based on secondary data enables the measurement of error alignment and the rate of compliance inhibition.This work defines a robust and ethical framework for use, compliant with the requirements of the EU AI Act, positioning the constrained persona as a filter for user resistance prior to in vivo clinical research.