Abstract This paper introduces the general idea of error contrasts, their properties, and uses. Error contrasts is a way for partitioning the error space and depend on it's canonical representation, that is the vector space of linear unbiased estimators of zero. It appears that breaking down sums of squares of error to contrasts each with one degrees of freedom is very useful in inference problems. In particular, negative variance values of ANOVA estimates, testing problems, and constructing exact confidence intervals for a positive linear combinations of variance component. Illustrations in unbalanced models in case of one way random analysis of variance and simple nested designs are presented.