How could records, evidence links and answer keys be delivered?

A release could deliver source records in JSONL and relationship tables in Parquet. Each record would carry a stable `record_id`, `record_type`, `case_id`, `schema_version` and payload. A relationship row could contain `from_id`, `to_id`, `relation_type` and `claim_id`.

Typed relationships such as `approves`, `executes`, `supersedes` and `supports` help customers reconstruct the evidence graph. Their semantics need documentation: an `approves` edge should identify the authorised decision, while a `supports` edge should identify the specific claim receiving support.

Task inputs should be separated from restricted evaluation material. A public task row could provide `task_id`, `visible_record_ids` and `requested_output_schema`. A restricted row could hold required claims, acceptable evidence paths and prohibited conclusions. Explicit partitioning helps resellers avoid including the answer key in an indexed customer corpus.

A validation report could check unique identifiers, valid edge endpoints and conformance to the declared schema. Those structural checks do not establish that every business relationship is plausible.

Licensed resellers could preserve these components as separate downloadable assets or service responses, with manifests documenting which parts customers receive for training and which remain evaluator-controlled.