arXiv:2609.17288v2 Announce Type: replace Abstract: Under non-informative complex survey designs, dependence among sampling indicators prevents the direct application of standard i.i.d. semiparametric efficiency theory. We introduce a \emph{reference Poisson experiment}, a hypothetical i.i.d.\ experiment constructed from the limiting inclusion probabilities of the original dependent design. Under suitable conditions, the actual observed-data experiment is in fact locally asymptotically normal, with the same tangent-space geometry as the reference Poisson experiment. Consequently, the efficient influence function derived by orthogonal projection in this hypothetical i.i.d.\ setting characterizes efficient estimators under the original dependent design. This yields a systematic approach to efficiency analysis for general semiparametric full-data models and pathwise-differentiable targets under complex survey designs. Under the joint superpopulation--design law, the efficiency bounds depend on the design only through its limiting inclusion probabilities. For finite-population targets whose first-order expansions are governed by the full-data efficient influence function, the same estimators are efficient under both superpopulation and finite-population centering, with bounds linked by a Pythagorean identity yielding a generalized finite-population correction. We establish efficiency of cross-fitted mean and conditional-mean regression estimators under explicit design-specific sufficient conditions. Simulations and California Academic Performance Index data illustrate the theory.