Empirical evaluation of instructional innovations traditionally relies on time-consuming and resource-intensive classroom trials, limiting the pace and breadth of design exploration. This research introduces a “classroom-free” simulation framework that models learner self-drive and instructional barriers via closed-form stochastic operators and integrates gradient-based optimisation with a large-language-model (LLM) feedback loop. Our contributions are threefold. First, we derive a mathematically grounded simulator whose self-drive component admits analytic gradients and whose barrier operator captures dynamic disruptions with Gaussian noise. Second, we embed an LLM-in-the-loop optimiser that generates interpretable micro-tweaks to instructional design parameters, enabling human–AI co-design. Third, we validate the simulator against a real-world dataset of 45 Grade-8 students’ pre–/post-test gains, demonstrating that our model predicts the direction of learning gains (Wilcoxon W = 173.5, p<0.001). Ablation studies reveal that single-shot LLM feedback accelerates early optimisation but plateaus without continuous prompting. We discuss limitations—Gaussian noise assumptions, learner homogeneity, and ceiling effects—and outline future work on heteroskedastic noise, hierarchical priors, and dynamic LLM prompting. Our results suggest that computational proxies can triage instructional designs rapidly and at scale, offering a practical virtual wind tunnel for educational research and policymaking.