Outside evaluators can help close the gap between AI capabilities and our understanding of them. But access, independence and meaningful consequences matter more than a benchmark score.