Smolcluster GRPO tests 64-token summaries

The core issue is reward mismatch: forcing exactly 64 tokens can help the task, but it also fights overlap metrics that already punish brevity, so the baseline can look worse than it is.

• –DeepEval plus a GPT-5 judge is the right call here because faithfulness and clarity matter more than n-gram overlap for summary quality.
• –The 3x Mac mini plus MLX/vLLM-metal setup is a credible low-cost RL lab for small-model experimentation, not just a hardware stunt.
• –If the next SFT/DPO run beats GRPO, that would suggest the optimization problem is simpler than the reward design implies.
• –The most interesting result will be whether length-conditioned supervision can hold the 64-token target without the metric collision seen in BLEU and ROUGE-L.