8GB of VRAM, 135M params, under 100 lines: enough to run SFT, DPO, and GRPO end to end. The useful result isn't the score, it's a cheap bench where SFT drifts more than on-policy RL. Post-training…