How can generated exports preserve precise evidence references?

A release can provide JSONL records for event-shaped data and Parquet tables for larger relational collections. Use separate identifiers for the business object, the evidence record and the file version. A case identifier alone cannot resolve a citation when several documents support the same case.

An evidence schema could contain `evidence_id`, `case_id`, `artifact_version_id`, `media_type` and `content_digest`. A locator table could add `locator_id`, `artifact_version_id`, `page_number`, `section_key` and `text_span`. For structured rows, use a row key instead of pretending that every source has pages.

Represent claim support through an edge table containing `claim_id`, `evidence_id`, `locator_id` and `support_relation`. Relations can distinguish direct support, qualification and contradiction. Require each locator to resolve against the cited version.

A loader can reject orphaned edges and mismatched digests before a task begins. An evaluation hook can then verify whether the agent's returned citations resolve to the intended passage, independently of whether its interpretation of that passage is correct.