Data quality control (QC) is one of the most critical components in data assimilation. It is not only necessary in order to warrant that good observations are assimilated, but also to ensure that the observations are consistent with what the assimilation system can resolve. The data QC procedure in a mesoscale NWP system is often developed solely to filter out the bad observations, and is typically ignored or with insufficient consideration for observation-model consistency in terms of the representativeness error, which can vary greatly with models and/or the model resolutions. The NCAR real-time four-dimensional data assimilation and forecasting (RTFDDA) system is a multi-scale rapid-cycling operational NWP system which ingests observations into the WRF model directly during the model integration. The direct ingestion of an observation into the full-physics WRF model renders the system more vulnerable to any problematic observations. To deal with the issue, a unique off-line data QC system, which assigns a QC confidence flag with an objective approach for each observation and also uses it to weight the observation in data assimilation, has been built upon the RTFDDA rapid-cycling forecasts (Liu et al. 2006). In this paper, a more advanced (and more efficient) on-line QC scheme is described. The on-line QC module employs the same algorithms in the previous off-line system, but it makes use of the observation-time model analysis as a reference instead of the 1–6 hour forecasts used in the off-line QC system. Numerical experiments are performed to study the impact of the new QC scheme and compare it with the old one with RTFDDA. Comparison of the scheme used with RTFDDA and the more advanced NCAR next-generation 4D-REKF (4-D relaxation ensemble Kalman filer) is also conducted. It is found that the new on-line QC scheme significantly improves analyses and short-term forecasts for both RTFDDA and 4D-REKF systems. Furthermore, this WRF-RTFDDA on-line QC scheme has been deployed by AirDat LLC and used as a key component in their real-time TAMDAR QC procedure.