Experimental Validation of Configurational AI Architectures

Configurable intelligent design is formalized here as the selection of a consistent configuration from interdependent alternatives under constraints of budget, quality, risk, and compatibility. It is examined not as a universal solver but as a class of domain-specific systems, each defined for a task family with explicitly formalizable variables, hard constraints, and objectives. Locally correct predictions for individual components do not guarantee a globally feasible solution. This study compares direct neural-network inference, the exact HIM-D solver, and a procedural hybrid whose network candidate is checked by an independent verifier, with an exact search invoked whenever a constraint is violated. A reproducible synthetic corpus of 24,000 problems across educational, engineering, and commercial scenarios was assessed for feasibility, optimality, robustness to distribution shift, repeatability, and latency. Pure neural-network models were feasible in only 17.40% and 19.12% of cases and degraded sharply out of distribution, whereas the hybrid maintained 100.00% feasibility and 83.08% joint optimality. A staged ablation identifies pairwise incompatibility density as the dominant source of that degradation, and constraint-aware correction alone raises feasibility without reaching a guarantee. The results therefore support separating probabilistic candidate generation from independent formal verification within a bounded class of specialized configuration problems, not a universal solver.