I found this surprising because that's such an old test that it must certainly be in the training data. I just tried to reproduce and I've been unable to get it

I found this surprising because that's such an old test that it must certainly be in the training data. I just tried to reproduce and I've been unable to get it (20B model, lowest "reasoning" budget) to fail that test (with a few different words).