The performance of model-based diagnostic techniques depends not only on the quality of the residuals generated using the models, but also on the method used to interpret the residuals. Robust residuals can often be interpreted deterministically, but noisy residuals can benefit from being interpreted probabilistically. A probabilistic framework enables the modeling of uncertainty and the relationship between multiple faults and multiple residuals. However, it is not well-suited for representing residual dynamics, and as a result, residuals must be assumed to not be autocorrelated. Since this condition is rarely met, this paper analyzes it to determine how residuals can be made to be fit the assumption, and the consequences when the assumption is violated. The paper demonstrates that fault probabilities determined using autocorrelated residuals are useful, but lack calibration. Two methods for removing autocorrelation are discussed and both are shown to result in probability estimates that trade refinement for calibration.