From Principles to Longitudinal AI Governance: An Evidence-Based Framework for Continuous Oversight

Organizations deploying AI systems face a governance challenge that existing frameworks do not fully address. Compliance with governance requirements at a given point in time does not guarantee that oversight practices will persist as systems are retrained, updated, and redeployed. This paper proposes the Framework for Institutional AI Responsibility (FIAR), an evidence-based framework for longitudinal AI governance that distinguishes transient compliance states from persistent governance capability. FIAR structures oversight through a three-layer architecture connecting normative principles, operational controls, and lifecycle evidence, and introduces an evidence pipeline that transforms technical and governance artifacts into indicators, signals, and institutional actions across successive system versions. The framework yields two complementary governance outputs: point-intime compliance assessments for individual system versions and longitudinal maturity evaluations that characterize governance capability across projects and successive system versions. The framework is illustrated through a case study in a public research institution developing healthcare AI in partnership with the national Ministry of Health. The case demonstrates how FIAR supports governance
auditability and structures fairness evidence for future longitudinal tracking, without requiring access to sensitive data or proprietary models.