SISTEMAS DE INTELIGÊNCIA ARTIFICIAL COM NÚCLEO IMUTÁVEL E SUPERVISÃO COGNITIVA INTERNA: UMA ARQUITETURA PARA AMBIENTES CRÍTICOS E REGULADOS

Abstract The widespread deployment of Large Language Models and automated decision-making systems in socially critical domains has exposed fundamental structural limitations in contemporary AI architectures. In particular, the incorporation of continuous adaptive mechanisms during operation introduces opacity, non-determinism, and the absence of formal safety guarantees, making such systems incompatible with environments that demand predictability, auditability, and accountability. This work proposes a governable AI architecture grounded in principles derived from high-reliability and safety-critical engineering. The core of the architecture is the strict separation between immutable system identity and operational context. An immutable core encapsulates validated parameters, constraints, and decision logic, while execution occurs within ephemeral sessions that leave no persistent state. Adaptation and learning are removed from runtime and confined to controlled offline processes. The architecture is complemented by three integrated pillars: controlled learning through offline retraining, structured memory for auditability and forensic analysis, and internal cognitive supervision responsible for validating decisions against explicit safety and governance constraints. Together, these mechanisms enable supervised decision cycles, deterministic reinitialization, and bounded evolution without compromising system integrity. By redefining intelligence as a property constrained by governance rather than adaptive autonomy, this work outlines a paradigm shift toward AI systems that prioritize safety, responsibility, and resilience over unchecked optimization. The proposed architecture provides a concrete technical foundation for deploying AI in regulated and mission-critical environments while preserving human accountability and systemic trust.