How We Caught an LLM Scoring Bug and Fixed It Without Compromising Our Research HackerNoon