<p>Manual sampling and heuristic tuning in high-throughput IC tray injection molding frequently delay the identification and correction of incremental process drift, leading to out-of-specification weight variation and increased rework risk. This study introduces an evidence-driven, event-triggered closed-loop decision-support architecture for real-time weight stabilization. The system integrates in-line per-part weighing with streaming machine signals and initiates parameter adaptation only when statistically sufficient evidence of persistent deviation is detected. A parallel Shewhart-CUSUM evidence gate converts continuous quality monitoring into discrete intervention events, enhancing intervention efficiency while maintaining stability during in-control operation. To eliminate on-machine trial-and-error, a data-driven surrogate environment is developed to approximate the nonlinear process-quality relationship within a bounded operating region. Within this environment, Actor-Critic reinforcement learning generates Top-K candidate recipe adjustments for engineer review under human-in-the-loop actuation. Surrogate-supported analyses of representative drift episodes indicated that the recommended parameter updates were associated with recentering of the weight distribution toward the target and reduced dispersion within the evaluated event windows. Furthermore, a traceable representative deployment validation episode based on an executed recommendation exhibited improved measured post-intervention outcomes, including lower mean error relative to the target, reduced standard deviation, and a substantially lower out-of-specification rate. The framework was deployed as a human-in-the-loop decision-support service on an active production line for six months. During this period, the average time required to complete a parameter adjustment following an abnormality decreased from approximately 10 to 7&#xa0;min, providing field-based operational evidence of improved workflow responsiveness under routine production variability. Overall, the proposed architecture establishes a governed and traceable workflow linking statistical process control (SPC)-based evidence gating, surrogate-mediated recommendation generation, bounded Top-K recommendation output, and engineer execution, thereby supporting an auditable and scalable decision-support mechanism for precision injection molding.</p>