The added sauce here is they're using it to bias the model during training, not just using steering vectors at inference time (though they do mention that). This is apparently effective at making the intended change in behavior without the lobotomizing side effects that steering vectors can have.
I've been referring to apparently this as "whatever a control vector is called in 2025" since they started doing it to dilute tokens under load: https://news.ycombinator.com/item?id=44082733
Tests are a way to write your logic twice (once for the code and once during the assertions) with the assumption that you're unlikely to make the same mistake twice.
Integration tests are better replaced by something like contract testing IMO to still retain the test parallelism.
