For the record, the TRACE output seems to reflect that the transformer is called from multiple threads but doesn’t seem to be an actual problem. I did hack a synchronize on a stack lock to help my other hack to save enhanced bytecode to my /tmp/jpa folder via https://github.com/scottmarlow/wildfly/tree/jpabytecode_to_disk and that seems to help avoid corrupting the saved (enhanced) Department.class file which has nothing to do with the actual TCK test failure. I attached the enhanced TCK test entity classes as shown by command javap -v -l -p -c -s -constants $CLASSFILENAME`