Groundedness checker

Score per-claim faithfulness against a context window.

Paste a context document and a candidate answer. The lab decomposes the answer into claims, then runs a Claude judge to mark each claim grounded / contradicted / unsupported, with evidence quotes.

Learning objectives

·Decompose answers into atomic claims.
·Distinguish grounded, contradicted, and unsupported claims.
·Build the per-claim citation pattern used in production RAG.

Inputs

Context (the retrieved documents)

Candidate answer

Default example contains a fabricated claim (the architect of the Duomo) and an unsupported number (population). A good groundedness judge should catch both.