GAIF: Governed AI Architecture Framework v1.0 - An AI-Native Enterprise Architecture Framework for Probabilistic, Learning, and Governed AI Systems

Enterprise architecture frameworks such as TOGAF, Zachman, and FEAF were designed for deterministic, request-response systems and do not address the unique architectural challenges of AI systems: probabilistic behavior, continuous learning after deployment, multi-agent emergent behavior, and regulatory complexity that requires a compliance architecture. This paper presents GAIF (Governed AI Architecture Framework), a comprehensive, AI-native enterprise architecture framework designed from the ground up for probabilistic, learning, and governed AI systems at enterprise scale. GAIF provides ten original contributions not found in any existing framework: (1) a formal governance tier classification algorithm; (2) a Governance Decay Rate (GDR) metric; (3) a Trust Inheritance Protocol for multi-agent systems; (4) a Behavioral Contract Architecture with dual-versioning; (5) a Composition Safety Degradation Model; (6) AI Architecture Fitness Functions; (7) an AI Bill of Materials Governance Architecture with tier-specific provenance requirements; (8) a Composition Cost Model for end-to-end cost attribution in multi-agent pipelines; (9) a Regulatory Adaptation Architecture for automated compliance propagation across AI systems; and (10) Continuous Architecture Assurance (CAA) with a computable AI Architecture Health Index (AAHI) that provides the first continuously computed, quantitative measure of enterprise AI governance health. The framework includes an eight-phase AI Architecture Development Method (AI-ADM), a four-layer AI Capability Reference Model covering 70+ capabilities, six safety architecture patterns validated against clinical data, a 25-question AI risk assessment framework, a five-level maturity model, and regulatory compliance mappings to HIPAA, the EU AI Act, and NIST AI RMF. GAIF has been developed and validated at a large not-for-profit health plan serving over six million members across Databricks, Azure AI Foundry, and Snowflake platforms. Contributions 1-6 are informed by real-world deployment experience and the author's independent research program (PHI-GUARD, ContamPerc, EMG). Contributions 7-10 are proposed architectural primitives with formal definitions, included to establish priority and enable independent evaluation. GAIF is offered as an open framework specification intended to serve as an AI-native alternative to TOGAF for the enterprise architecture and AI governance communities.