Ground Truth
The label you treat as correct when training and scoring a model — your best available answer, not necessarily the real one.
Think of It Like This
Like the answer key in the teacher's grading manual: you treat it as absolute truth for grading, even if question 4 had a typo.
The name oversells it, and that's worth holding onto. Ground truth is a record of what somebody decided, or what some upstream system logged, and it carries every mistake and ambiguity that process contained.
Which makes it a ceiling rather than an input. A supervised model approximates its labels, so if two annotators disagree on a fifth of your examples, nothing you train can be more accurate than that disagreement — and measuring it costs a few hundred double-labelled rows.
The other reason to stay sceptical: a label derived from an outcome can leak into your features through a column that describes the same event. When a feature predicts the label suspiciously well, check which one came first.