Comment by xpct

It's probably fuzzily fixable by including instruction authority levels in the training data. Can't expect much more than that, given that the model itself is fuzzy.