Concepts & definitions
What is evidence coverage in legal AI?
Definition
Evidence coverage describes which authorized sources and passages were available, retrieved and actually examined for a legal AI task. A complete document inventory does not prove that every page was extracted, and a complete retrieval scan does not prove that every passage was analyzed. Coverage claims should identify the task boundary and material gaps so a selected answer is not mistaken for an exhaustive review.
Four different boundaries
Storage means the file exists. Extraction means readable text or other supported evidence was produced. Retrieval selects evidence for the task. Analysis means the selected evidence was actually considered. Failure at one boundary should not be hidden by success at another.
A caption-only extraction may create a searchable source while missing a scanned body. An OCR review must therefore check substantive pages, not merely whether some text was found.
A practical example
A request to identify every dated event in a bundle cannot be satisfied by searching for the word chronology and reading ten high-scoring passages. A relevant date could sit in a late exhibit with no matching keyword.
An exhaustive review should account for the eligible corpus, reconcile findings and identify evidence gaps. A normal question can use targeted retrieval without claiming exhaustive coverage.
What to ask during evaluation
Place a decisive event on a late page and contradictory evidence in a later source. Test whether the answer finds both, cites the right passages and distinguishes unavailable pages from optional structured enrichment that is still pending.
A result should say what was examined and what remains uncertain. Empty findings after a provider failure are not evidence that no contradiction exists.