arXiv:2604.22902v1 Announce Type: cross Abstract: We formalize a new inference problem: requirement-steered interface type inference. Given spatiotemporal observations of a physical system and functional requirements, the task is to infer what kind of interface must separate the system's interior from its environment for those requirements to be satisfiable. Unlike classical constrained design, which optimizes parameters within a pre-specified object type, here the type itself is unknown. We cast the problem as constrained variational Bayesian inference over Markov blanket partitions and introduce Constrained Dynamic Markov Blanket Detection (C-DMBD). The algorithm extends DMBD by steering blanket discovery toward functional targets through Lagrange multipliers updated by dual ascent. These multipliers penalize violations computed from model-predicted blanket statistics, allowing requirements to shape both the inferred partition and the interface dynamics. The framework yields three phenomena unavailable to classical design: intra-family navigation, where one interface type supports different functional modes; family transition, where changing requirements induce a discontinuous shift in interface type; and ontological disambiguation, where requirements resolve ambiguities left open by physical data alone. The converged multipliers form a certificate of functional effort, recording which physical properties the inferred interface resists satisfying. Finally, we argue that the generative model family associated with a cup's Markov blanket belongs to the designer, not to the cup. The cup is inert; the model family is the designer's representation of the dynamics its surface can sustain. This yields a loop in which the designer encodes a cup-user model into the surface, the user reconstructs it through active inference, and physiological data reveal the gap between the designer's prior and the user's actual generative model.