Genetic Algorithms (GAs) are promising for receding-horizon control in nonlinear, constrained systems, but their wall-clock cost is dominated by objective evaluations. We propose a rejection-based GA tailored to horizon-structured genomes that performs exactly one plant-evaluation batch per generation and filters unevaluated offspring through a lightweight, calibrated classifier. On a standard nonisothermal Continuous Stirred Tank Reactor (CSTR) benchmark, we compare four optimizers under identical settings: a baseline GA, our committee-gated RB-GA, the baseline with Simple Variable Population Sizing (SVPS), and the recently proposed Two-Level Adaptive Simulated Binary Crossover (TLASBX). RB-GA consistently improves tracking accuracy over the baseline at lower generation budgets. Baseline+SVPS and TLASBX closely match the baseline. We attribute this outcome to the limited number of generations, which constrains their learning effects. Overall, the results indicate that the proposed mechanism enhances optimization convergence in expensive-evaluation regimes.