Twice the data, less power: my stability rule got blinder the harder I looked

I was comparing two models on a frozen 159 task suite and I had a problem most eval code does not admit to having: the same model, asked the same question three times, sometimes di