RooFit's caching can lead to wrong results when batch computations used.

When a PDF doesn't implement the faster batch interface, RooFit's old, single-value computations have to be used as a fallback. If RooFit, however, tries to precalculate those values, the nodes of the computation graph will always yield the same wrong value, since they are switched to "always clean".

This e.g. happens when a node of the graph doesn't depend on parameters, but only on observables.