It trains a small decision head onto a frozen model. Base weights unchanged, served with plain vllm serve. Across 6 open models (Qwen3.5, Qwen3, Gemma 4): • thinking off: +8 to 13 points of accuracy…