Structural Limits of Probabilistic AI Architectures under the EU Artificial Intelligence Act

This paper presents a system-level feasibility analysis of probabilistic artificial intelligence architectures under the regulatory requirements of the European Union Artificial Intelligence Act. Rather than evaluating documentation practices or provider-specific implementations, the analysis addresses a logically prior question: whether certain classes of non-deterministic, sampling-based AI systems are, in principle, capable of fulfilling the material obligations imposed by the regulation. The paper demonstrates that, under sustained regulatory-style interrogation, probabilistic architectures may enter a regime of architectural incoherence in which traceability, stable measurability, causal attribution, and effective governance cannot be reliably maintained. In such cases, compliance artifacts may remain formally valid while becoming materially ineffective. The analysis is strictly structural and conditional. It does not constitute a legal opinion, a product audit, or an assessment of any specific proprietary system. All claims are derived from explicitly stated premises and empirical system behavior. An audit marker based on a cryptographic hash is provided to ensure document integrity and reproducibility.