Synthetic patient records are increasingly used where real clinical data are scarce, restricted by privacy rules, or expensive to label. How far performance on such data predicts performance on genuine records remains unclear for tabular cardiac prediction. This study fixes the architecture and varies only the data. One feed-forward network, unchanged in structure, was trained on a synthetic corpus of 9,000 patient records and on the real UCI Cleveland Heart Disease dataset of 303 patients, under identical preprocessing, class-weighted optimization, and evaluation, so that any difference reflects data provenance rather than model design. On the synthetic data the network reached an accuracy of 0.869 and an AUC of 0.940. Transferred to the real data, accuracy dropped to 0.787 while AUC stayed at 0.920: the model still ranked patients well, but its calibrated classification deteriorated. A second experiment on the real data added early stopping, five-fold stratified cross-validation, a hyperparameter grid search and a fold-level ensemble. Accuracy returned to 0.869 and AUC rose to 0.953, in line with published results on the same benchmark. The search selected the most strongly regularized configuration, which points to overfitting on a small sample rather than any intrinsic difficulty of the real data. Two practical implications follow. Synthetic data can make a classification task look easier than it is, and the gap this creates is largely closed by standard regularization. Synthetic datasets are suitable for model prototyping; however, conclusions derived from them must be validated using real clinical data before clinical interpretation.