For example, the draftquality models require non-public, deleted article content. We shouldn't be copying that to labs boxes.
One workaround might be to calculate features on our development machine, then export just the evaluated feature values to ores-compute, omitting the article text.
A longer-term solution would be to secure the training compute box, but that would have to be in the production cluster. I'm not sure if the security tradeoff makes sense there.