Ah after reading your text I understand - you do not write anything about your system (which database, openLCA version, impact method, and also not how many iterations you have performed) - we changed the uncertainty in ecoinvent with the latest version, changed also the monte carlo simulation in 2.0.2, and will release finally tomorrow the new method package that has impact directions for LCIA methods that will make the impact calculation "save" (in the sense that an elementary flow is not put on the wrong side, i.e. a resource is not considered as emission and vice versa). You could check if you have a similar picture for inventory flows (if not it is probably the impact method) and check if the number of simulations is high enough (since in the end, MC results are just random, and if you have 5 iterations or so then results may be quite off, just by chance).
All the best,
Andreas
