Which relationships can a generated evidence dataset represent?

A release can include a relationship graph connecting the question under review to the material needed to answer it. Useful entities include review claims, control requirements, business events, records and evidence bundles. The graph describes how those entities relate, rather than treating every attachment as equally supportive.

An expense review claim may depend on both an eligible purchase and an independent approval. A receipt supports the purchase details. An approval record supports authorisation. A role assignment establishes whether the approving role had the required authority. No single document establishes the whole claim.

Generated relationships can express `supports`, `contradicts`, `depends_on` and `covers_part_of`. They can also identify the exact passage or structured field on which a relationship rests. A buyer can ask for reasoning annotations separately from the documents exposed to the model.

This structure supports tasks that require a minimal sufficient bundle, a gap explanation or a defence of a particular inclusion. It also makes partial evidence meaningful: a record can be useful without being enough, and a bundle can support one part of a claim while leaving another unresolved.