Regulatory Learning or Precaution? Capital, Reporting, and Supervisory Architecture Under Ambiguity

This study constructs a two-period model where a regulator determines the level of precautionary capital, information-generating reporting, and the design of supervisory information systems, while providers respond by participating. When reporting successfully produces a verified diagnostic, continuation capital is set afterward, but reporting incurs various costs, including variable, participation, and fixed setup costs, before its information is utilized. In a Bayesian context, the diagnostic's gross decision value is weakly nonnegative because it can be disregarded. Reporting is only activated if its discounted decision value and any screening benefits outweigh its net costs and setup expenses. Under recursive maxmin assumptions, an admissible prior that makes the adverse state certain introduces a certainty boundary: diagnosis cannot alter the continuation capital once the conditions are met, and precaution substitutes for learning. Priors that are uniformly interior allow for ongoing learning. Recursive smooth ambiguity models positive reporting at finite levels of ambiguity aversion and converges to the maxmin boundary under specified conditions, without implying overall monotonicity. A joint-selection theorem compares scenarios with no reporting, common, and specialized reporting architectures after optimizing intensity. An architecture that keeps the common experiment and adds an ignorable signal slightly improves gross information but may reduce net surplus. Kenya’s virtual-asset framework provides a dated institutional example but neither calibrates the model nor reveals its core mechanism. The results are supported by analytical proofs, independent recalculations, and reproducible code.