Machine learning models are increasingly used for encrypted traffic analysis because payload inspection is limited. However, high internal detection accuracy does not necessarily indicate that model outputs are reliable for forensic interpretation. This study examines this issue through cross-dataset false positive analysis. Four machine learning models are trained and evaluated internally on the CIRA-CIC-DoHBrw-2020 dataset. Random Forest achieved the highest internal accuracy of 99.62\%, while CatBoost achieved 98.92\% and was used for detailed cross-dataset and explainable analysis. The trained CatBoost model was then applied without retraining to 947,997 benign flows from the CESNET-TLS-Year22 dataset, where 41,238 flows were classified as malicious, resulting in a false positive rate of 4.35\%. This result shows that strong internal performance can still produce misleading alerts when the traffic environment changes. SHAP and permutation importance are used to examine feature dependence in the model predictions. Both methods identify a consistent set of aligned flow-level features, with packet size mean showing the strongest influence in the SHAP analysis. These features support classification but are not unique to malicious traffic and can also occur in benign encrypted communication. Therefore, feature importance alone is insufficient for forensic interpretation. The findings indicate that machine learning alerts should be treated as investigative leads rather than direct forensic evidence, and false positive analysis should be considered when interpreting model explanations across different traffic environments.