CASE Framework Maps Agentic AI Governance to Four Classical Sciences

The August 2026 arXiv preprint formalizes the CASE architecture as four non-compositional layers. Control theory treats intent as setpoint with guardrails as feedback. Complex adaptive systems theory addresses emergence that breaks single-agent verification. Supervisory cybernetics applies the Law of Requisite Variety to show unaided human oversight fails structurally. Engineering operations extends error budgets to decision quality. Cross-layer coupling conditions include a zero-touch deployment paradox that improves one layer while degrading others.

Empirical sections report three findings. 82 percent of documented production agent failures follow multi-layer trajectories. None of 22 ecosystem tools covers full emergence-layer assurance. All 35 scored public deployments sit in the lowest maturity band. These data define the Emergence Gap between realized risk at the collective layer and capability actually deployed.

The five-level maturity model uses a non-compensatory bottleneck-weighted index. It directly operationalizes EU AI Act Article 14 requirements for effective human oversight. Only architectures that satisfy requisite variety can convert oversight from ceremonial to functional. The model supplies an assessment instrument grounded in production enterprise platforms rather than process checklists.