Detect and close ground-truth leakage in LLM eval datasets — cases where the model under evaluation can see material that contains or strongly implies the expected answer, so eval scores are inflated and no longer reflect real capability. Use when building, reviewing, or debugging an eval/benchmark dataset (especially for RAG or tool-using/agentic systems), when eval scores look suspiciously high, when a weak or blind baseline still passes, or when deciding whether retrieved context, tool/observation outputs, dataset fields, or upstream-stage outputs leak the target answer.
2026-05-31