A recent formalism recasts the modern compiler as a distributed optimization system, mapping compiler architectures onto decomposition strategies from Network Utility Maximization (NUM). Under this framing, a JIT compiler performing profile-guided optimization implements dual decomposition: a runtime master broadcasts shadow prices (execution frequency data), and the optimizer allocates its budget where the price is highest. We test this hypothesis empirically by instrumenting MIR, a lightweight JIT compiler infrastructure, with a shadow-price-responsive inlining threshold. Across 3,000 inlining decisions (750 per condition x 4 conditions), we find strong monotonic correlation between shadow price and inlining outcome (Spearman rho = 0.71, p < 10^-114), large effect sizes (Cohen's d = 2.00), and 98.4% decision agreement under plus-or-minus 20% price perturbation. Stratified analysis reveals that shadow-price sensitivity is optimization-level-dependent: conservative settings (O0) show reduced responsiveness (rho = 0.50) due to threshold ceiling effects, while moderate and aggressive settings (O1, O2) show strong responsiveness (rho = 0.82). Size-stratified analysis identifies a decision boundary at 100 to 200 IR instructions where shadow prices govern the inlining outcome. These results corroborate the NUM formalism as a predictive framework for compiler inlining behavior and establish experimental infrastructure for further validation. Version 1.1 incorporates expanded discussion of the scope of claims, addressing underdetermination and the epistemological status of the standalone decision function as a passed falsification test. Design implications are strengthened with specific falsifiable predictions including O(1/sqrt(k)) convergence bounds for cross-module price propagation and O(log n) stability bounds for adaptive inlining. This preprint is archived for open access and to support reproducibility
compiler optimization · JIT compilation · network utility maximization · dual decomposition · profile guided optimization · inlining · empirical validation · shadow prices · resource allocation · MIR compiler