Addressing the challenges of opaque audience voting data and variable elimination rules in competitive reality shows, this study constructs a quantitative framework integrating latent variable reconstruction, rule auditing, and dynamic optimization. First, a Bayesian state space model is employed to perform latent state reconstruction on audience support rates across thirty-four seasons. Using Laplace approximation techniques to quantify estimation uncertainty, the analysis reveals significantly heightened sensitivity of voting fluctuations to late-stage decision-making. Subsequently, counterfactual simulations under different scoring rules were conducted using the reconstructed data, deeply analyzing the systemic differences between percentage-based and ranking-based systems in addressing expert bias versus public popularity bias. Furthermore, a panel regression model was applied to deconstruct the influence mechanisms of factors like professional dancer background, contestant occupation, and age on advancement probabilities, identifying significant divergences in the driving logic between expert judging and public voting. Finally, addressing the vulnerability of existing mechanisms to noisy data, we propose an uncertainty-aware weighted moving average mechanism. By dynamically adjusting voting weights, this approach simultaneously enhances the robustness and fairness of elimination decisions. Experiments demonstrate that this framework effectively restores latent competitive dynamics, providing a robust algorithmic foundation for optimizing dual-track judging systems.

bayesian inference, counterfactual simulation, dynamic weight allocation