This article addresses the performance of Dempster-Shafer (DS) theory, when it is slightly modified to prevent it from becoming too certain of its decision upon accumulation of supporting evidence. Since this is done by requiring that the ignorance never becomes too small, one can refer to this variant of DS theory as Thresholded-DS. In doing so, one ensures that DS can respond quickly to a consistent change in the evidence that it fuses. Only realistic data is fused, where realism is discussed in terms of data certainty and data accuracy, thereby avoiding Zadeh’s paradox. Performance measures of Thresholded-DS are provided for various thresholds in terms of sensor data certainty and fusion accuracy to help designers assess beforehand, by varying the threshold appropriately, the achievable performance in terms of the estimated certainty, and accuracy of the data that must be fused. The performance measures are twofold, first in terms of stability when fused data are consistent, and second in terms of the latency in the response time when an abrupt change occurs in the data to be fused. These two performance measures must be traded off against each other, which is the reason why the performance curves will be very helpful for designers of multi-source information fusion systems using Thresholded-DS.