Ground truth is the known-correct answer for a task, used as the reference when scoring output. When ground truth exists, evaluation can be reference-based and objective; when it does not, you rely on reference-free judges or rubrics. Ground truth quality sets the ceiling for evaluation quality.
Why it matters
When the correct answer is known, evaluation can be objective and automatic. Ground truth quality therefore sets the ceiling on how much any score can be trusted.
How it works
Reference-based evaluation compares output to the known answer; reference-free evaluation relies on judges and rubrics when no single answer exists. Ground truth is usually built by human annotation and promoted into golden datasets.