> No harness at all and very simplistic prompt

TBF, that's basically what the kaggle competition is for. Take whatever they do, plug in a SotA LLM and it should do better than whatever people can do with limited GPUs and open models.