With the support for iterations/invocations (#68), do we need to rethink handling of faulty runs? We might want to keep for each result (data point/measurement) whether it was recorded from a faulty run.
