How should the data model represent the purpose of an evidence package?

Represent the review as linked entities rather than a folder with a single completion label. A compact schema can contain `review_id`, `claim_id`, `process_step_id`, `record_id` and `relationship_type`. Claims describe what must be established; records hold the generated material an agent can inspect.

Relationship types can distinguish `supports`, `contradicts`, `qualifies` and `requires_follow_up`. An approval record may support an authorisation claim while a cancellation notice qualifies its continuing validity. Preserving both relationships prevents a file’s presence from becoming an automatic pass.

For an evaluation interface, expose actions such as `list_claims`, `read_record` and `submit_evidence_map`. The submitted map can contain a conclusion for each claim plus the record identifiers and passages used to justify it. Evaluator-only relationships can remain outside the agent’s accessible store.

JSONL suits individual entities and edges; Parquet can support analysis across submissions. Require stable identifiers across both representations. The important integration property is traceability: a reviewer should be able to travel from a conclusion to a cited record and back to the exact process claim it is meant to establish.