I hope you read the disclaimer that is printed when you launch JAGS – you know the one about ABSOLUTELY NO WARRANTY – because it means what it says.
JAGS 1.0.3, which has just been released, fixes a serious bug in the probit link function, and its inverse, the phi function. These functions returned results based on the upper tail, not the lower tail, of the normal distribution function. Consequently, all the coefficients in probit regression models have the wrong sign. Thanks to Richard Morey for sending me a simple example which showed up the problem.
Obviously, this bug should have been found and fixed well before version 1.0, and I am very embarassed to be revealing it now. Unfortunately, the incorrect results had become entrenched in the benchmarks and so were self-perpetuating. A new release of the classic bugs examples includes updated benchmarks for the BEETLES and LSAT examples.