Ground Truth

The reference used to determine the expected outcome of an evaluation case. It may come from a deterministic rule, a system of record, a normative document, or expert annotation. Its quality depends on provenance, annotation instructions, coverage, and versioning; a label without that context merely moves uncertainty into the dataset.

Open-ended tasks often allow several acceptable answers or legitimate disagreement among evaluators. In these cases, the reference can record criteria, required evidence, and a distribution of judgments rather than one sentence. Changes to ground truth need review and history because they retroactively change the meaning of the metrics.