From Finite Evidence to Experiment Choice: Coverage-Controlled Admissible Worlds for Cost-Aware Decision Resolution

A difficult experimental question often appears one step before experiment selection: not which experiment is best, but which worlds are still admissible after the evidence already collected. We develop an evidence-first interface between finite-sample statistical uncertainty and cost-aware experimental planning. Finite noisy observations are inverted into a coverage-controlled set of admissible worlds, which is then mapped through a downstream decision rule. If all admissible worlds agree, acquisition stops. If they disagree, a target-aware planner searches for the least-cost admissible experiment satisfying a declared separation requirement; if no such experiment exists under the available family, margin, or budget, the procedure returns a typed obstruction rather than an unsupported decision. In exact Bernoulli validation, nominal 0.95 admission coverage was preserved across 999 parameter values and sample sizes 10–200, with minimum observed coverage 0.950009652. For evidence 17/20, the admissible grid spans 0.63–0.95 around a 0.80 deployment threshold: an existential contradiction can be exposed with eight additional observations, whereas the closest finite-grid boundary pair requires 10,705. A controlled ML deployment example yields a worst-case global certification cost of 112,402.5. An exact target-aware dynamic-programming baseline attains the same cost, showing that the finite planning layer is not the contribution. The contribution is the statistically auditable bridge from noisy evidence to admissible worlds, decision disagreement, and explicit resolution or obstruction.