Toward Trustworthy LLM Support for Foster Families: Architecture and Formative Evidence from an Early-Stage System

Foster families routinely navigate intertwined relational, behavioral, educational, administrative, and health-related questions, often when professional help is not immediately accessible. Large language models (LLMs) could lower the threshold for obtaining orientation and emotional support, but an unconstrained assistant may produce unsupported or overly authoritative guidance in a setting involving vulnerable children and caregivers. We present Nesirotstvo, an early-stage Russian-language support system designed as a complement, not an alternative, to qualified services. Its orchestration layer maintains bounded dialogue context, selects among history-, memory-, and retrieval-based response routes, searches a privately maintained knowledge base before selectively consulting allowlisted web resources, and applies a conservative policy when evidence is absent. We also report an author-led, LLM-assisted descriptive audit of 65 exchanges across 18 recorded pilot sessions. Seventy-five percent of exchanges had attached sources, while the audit also exposed long responses, substantial tail latency, and interactions without retrieved evidence. These observations demonstrate implementation feasibility and inform design priorities; they do not establish factual accuracy, safety, user satisfaction, or family outcomes.