LEKIA: Expert-Aligned AI Behavior Design for High-Risk Human-AI Interactions

Large language models (LLMs) have demonstrated technical accuracy in high-risk domains, such as mental health support and special education. However, they often fail to meet the nuanced behavioral expectations of domain experts. This gap constrains AI deployment in sensitive settings. To address this challenge, we introduce LEKIA (Layered Expert Knowledge Injection Architecture), a novel framework built upon the principle of expert-owned AI behavior design. LEKIA's core innovation lies in its dual architecture: a three-layer knowledge injection system featuring our "Supervision Metaphor Cycle", and a dual-agent safety system ensuring robustness and consistency. We implemented and evaluated LEKIA within psychological support scenarios in special education. Experiments indicate that LEKIA improves performance by 14.8% over baseline, driven by substantive increase in alignment with expert expectations while preserving technical accuracy. Beyond providing a reproducible technical framework, this work demonstrates expert-expectation alignment as a measurable evaluation criterion with implications for AI deployment in high-risk domains.