From Architecture to Ecology: A Layered Framework for Governing Adaptive AI Systems

Pre-deployment specifications and evaluations remain necessary for AI assurance, but do not by themselves establish safe behavior across changing deployment conditions. We distinguish three layers of epistemic access to an AI system: an engineering layer (what we build), comprising specified artifacts and procedures; a capability layer (what we observe), comprising empirically assessed behavior; and a mechanism layer (what we understand), comprising internal computations whose causal interpretation remains incomplete. We argue that conflating these layers can create unjustified transfers of assurance, such as treating tested refusal behavior as a hard access boundary. Adaptive, tool-using, and continually updated deployments make these evidentiary distinctions especially consequential. We propose an ecological perspective that complements behavioral specifications with continuous, accountable design of the conditions that constrain and select behavior. We organize the governance environment into seven interacting pillars — Boundary, Rules, Incentives, Tolerance, Redundancy, Monitoring, and Exit — and cross them with the three layers to form a 3 × 7 governance matrix. Six design principles address model-independent boundaries, compensation for uncertain evidence, common-mode failures in redundant defenses, plural incentives, budgeted tolerance, and closed-loop revision. A hypothetical continual-learning design illustrates the matrix's constructive use; a source-bounded analysis of Moffatt v. Air Canada illustrates diagnostic questions without inferring undocumented implementation details. We map the pillars to NIST AI RMF 1.0, Regulation (EU) 2024/1689, and ISO/IEC 42001:2023, recognizing that these instruments already support lifecycle governance. The contribution is an integrative diagnostic framework, not a completeness theorem, a compliance certification, or an empirically validated safety guarantee.