For any linear model there is, according to an invariance argument, a uniquely defined prior representing ignorance, which will be called the inner prior. It is shown that the corresponding posterior probabilities are weighted averages of conditional confidence levels. This frequency interpretation compares favorably with the usual interpretation of confidence levels as unconditional probabilities of coverage. Of particular interest to users of statistical methods is that inner posterior intervals are always shorter than the corresponding confidence intervals and, sometimes, dramatically so.