Terahertz (THz) measurement-based parameter reconstruction and defect identification can support integrity assessment of multilayer composites in safety-critical structures. Reliable THz inversion nevertheless remains difficult because multiple physical parameters are usually strongly coupled and the reconstructed quantities are uncertain. This study proposes an uncertainty-aware physics-informed decoupled inversion method. A differentiable electromagnetic model constrains hybrid neural and parametric representations through measured responses. Sensitivity analysis guides parameter grouping and staged inversion, with larger weights assigned to frequencies that better discriminate the active parameter group. Conditional Bayesian inference propagates upstream uncertainty, while reliability-based covariance adjustment regulates its influence on later stages. A matched-model four-layer reflection benchmark demonstrates accurate material and geometry reconstruction and localization of prescribed defects. Evaluation on measured polymer-pellet transmission data further validates thickness reconstruction, with all holdout reference thicknesses contained within the adjusted nominal 95% credible intervals. The method provides physics-consistent parameter estimates, uncertainty characterization, and probabilistic defect decisions for coupled THz inverse problems.