The development of large-scale multiphysics simulations often comes with a series of decisions around the best physics representations and parameters for each sub-model configuration, with available options differing in physical assumptions, fidelity, and computational cost. Exascale computing has enabled large-scale multiphysics simulations by allowing expanded integration of complex multiphysics couplings and increasing fidelity in reproducing realistic physical and engineering systems. Exhaustive construction, calibration, and validation of all sub-model variants is still infeasible and ad hoc reliance on expert judgment risks discarding models with superior predictive capability for certain quantities of interest (QoIs). Choosing the correct sub-model requires a principled strategy that systematically prioritizes the sub-models that together result in the most valid simulation. This presentation introduces an iterative, uncertainty-guided framework for multiphysics simulation code development that integrates sub-model refinement with targeted data acquisition to maximize QoI-specific predictive accuracy under constrained resources. The framework treats the multiphysics system as a Bayesian network, encoding causal dependencies among sub-models and enabling propagation and aggregation of uncertainty to the QoIs. A sequential decision strategy couples Bayesian calibration, plausibility assessment, and validation under uncertainty to identify the sub-models with the most predictive configuration. The method is demonstrated in two representative applications: the progressive development of constitutive models for elastic–damage response of porous materials, and the 2D combustion in hybrid rocket systems with solid fuel ablation involving coupled atomization, gasification, and chemical-kinetics sub-models.