Designing an LLM Evaluation That Actually Tells You Something

Task-grounded eval sets, grader choice, sampling variance and the contamination traps that make benchmark scores misleading.