Three experiments investigated the properties of the time to determine confidence to determine the processing locus for the judgment of confidence. Results suggest that when the primary decision is made under speed stress, confidence is determined postdecisionally and involves a memory-based, computational algorithm. This strategy frees the primary decision of processing time and permits the accurate diagnosis of decision errors. When the primary decision is made under accuracy stress, however, the determination of confidence is initiated, or can even be completed, during the primary decision process. This strategy permits confidence to be used in the adaptive regulation of the decisional parameters during the decision process but yields poorer diagnosticity of errors when they occur. The latter finding also implies that primary decision latencies include time to determine confidence, rendering such data difficult, if not impossible, to model empirically. Implications for contemporary decision models that provide a basis for confidence in human judgment are discussed.