Human performance reliability for tasks in the time-space continuous domain is defined and a general mathematical model presented. The human performance measurement terms time-to-error and time-to-error-correction are defined. The model and measurement terms are tested using laboratory vigilance and manual control tasks. Error and error-correction data are ordered and the underlying density functions isolated. The Weibull distribution is best fit for time-to-first-error data, and the Log-Normal distribution is best fit for time-between-errors and time-to-error-correction data. The Normal distribution is rejected in all cases. Distribution parameter values are applied to the general mathematical model, and prediction made of human performance reliability for the tasks. It is also shown that task performance reliability improves with training on the tasks. (Author)