AIRA-F: AI-Native Reliability Architecture Framework for LLM-Powered Enterprise Systems

Enterprise deployments of Large Language Model (LLM)-powered systems introduce classes of runtime failure including hallucination, context overflow, retrieval staleness, agent retry loops, semantic drift, and governance bypass that traditional software architecture methods do not model. Existing frameworks such as the Twelve-Factor App, microservices reliability patterns, and MLOps maturity models address overlapping but distinct concerns and do not collectively cover the reliability challenges unique to LLM-powered systems. This paper proposes AIRA-F, the AI-Native Reliability Architecture Framework, developed through a practitioner-informed synthesis of recurring production LLM failure patterns and a review of software architecture, reliability engineering, and LLM systems literature. The framework applies a layered decomposition approach to model reliability concerns at each architectural boundary. The result is a six-layer taxonomy comprising Interaction, Context, Model Behavior, Tool/Agent Execution, Observability, and Governance, accompanied by a 13-failure-mode catalogue, per-layer metrics and controls, a five-level production readiness maturity model, and a reference architecture. AIRA-F is intended to provide software architects and engineering teams a practitioner-ready framework for designing, validating, observing, and governing production-grade LLM-powered enterprise systems. Empirical validation at scale is identified as a direction for future work.