Retract evidence, is an option that allows predictions to be made on variables that have evidence assigned.
Also see Retracted analysis.
If retract evidence is not turned on, and a variable that has evidence is queried (predicted), then the evidence value is returned. Often however, you want to predict each variable as if its own evidence was not present, which can be achieved by turning on the retract evidence option.
Consider the prediction of variables X, Y and Z, where all variables have evidence (all have data mapped to them). If the Retract button is checked, then the prediction of X will ignore any evidence set on X, using only evidence set on Y and Z. The prediction of Y will only use evidence set on X and Z, and the prediction of Z will only use evidence set on X and Y.
If you are querying (predicting) a joint distribution, evidence is retracted from all variables within the joint distribution.
If evidence has been retracted on a query distribution, the likelihood/log-likelihood associated with the evidence used to calculate the query distribution, will not equal the likelihood/log-likelihood) for all the evidence. (The Bayes Server API allows log-likelihoods to be returned for each query distribution.)