Atomic Facts

Self-contained units of information extracted from longer text, each representing a single verifiable claim or observation. In AI reliability research, decomposing model responses into atomic facts enables systematic comparison of what different models agree or disagree about. For example, a paragraph-long image description can be broken into individual facts like "the chair is red," "there is a table in the center," and "the room has wooden flooring." These atomic facts can then be clustered, compared across models, and annotated with agreement levels, enabling fine-grained reliability assessment that would be difficult with full-response comparison.