Detailed comparison between the predicted and true values for the three atmospheric parameters in the second-stage model corresponding to the case of R_p ∈ [0.8, 1.0) R_J. Each panel plots the model prediction (y-axis) against the true value (x-axis), with the dashed line marking the ideal 1:1 relation. The three columns correspond to the training, validation, and test datasets. The top row shows results for ﹩\mathrm{log}({T}_{p})﹩, where the predictions closely follow the unity line and retain strong performance even on the test set. The middle row presents predictions for ﹩\mathrm{log}({X}_{{\rm{TiO}}})﹩, which display higher scatter but still preserve a clear linear trend between predicted and true values. The bottom row shows the results for ﹩\mathrm{log}({X}_{{\rm{V\; O}}})﹩, with behavior similar to that of TiO. Across all parameters, the model exhibits the expected decrease in accuracy from training to test data, yet maintains a consistent correlation that demonstrates its ability to recover the main atmospheric trends from broadband photometry.