Our colleagues on HSC have reported a non-deterministic astrometric solution being generated by the stack. That's clearly a bug, and will have to be fixed (and it might be HSC specific). However, it's notable that this varies with the number of cores being used to execute the stack. That's hard for developers to spot when running tests on their own systems, but it would be easy to catch by varying the (virtualized?) hardware environment when running automatic regression tests.