LoRA adapters trained for 5 progressively shorter chain-of-thought styles on GSM8K, plus the eval artifacts behind the Pareto curve.