The behavior of a probabilistic spatial QC system as applied to daily temperature observations in Oregon is discussed. Examples are presented of how the system recognizes good and bad data in a spatial context. The confidence of an observation is determined through various statistics that are designed to assess how consistent the observation is with its neighbors. Two classes of consistency are assessed: (1) how well the observation can be estimated in its absence; and (2) how good the estimated observation for other, nearby stations is when the observation is used as one of the predictor values in a spatial estimate. The system systematically lowers the confidence of inconsistent data and raises the confidence of consistent data through a series of iterations. During each iteration cycle, model performance statistics for each day are compared against summary statistics describing the expected mean and variance of model performance for the same time of year. Daily statistics that deviate strongly from the expected distribution are assigned low confidence values, while those that fall within the expected range receive high values. These confidence values are used as PRISM weights in the next iteration, so that poor data have successively less influence and good data successively more influence on the system. The run ends once the confidence values have settled into equilibrium values. Particularly difficult situations that challenge the spatial QC system are presented, such a situation in which there are several bad observations within a small area on a given day.