Which fields make generated evidence reusable across environment components?

A release can define an evidence item with `record_id`, `record_type`, `source_system`, `version_id`, `content_ref` and `content_digest`. A separate anchor table can identify a paragraph, table row or structured field using `anchor_id`, `record_id`, `version_id` and `locator`.

Evidence links can then reference anchors rather than an entire file. A link might contain `assertion_id`, `anchor_id`, `relationship_type` and `agent_rationale`. This allows the environment to show exactly which passage the agent relied on and whether that passage belongs to the cited version.

JSONL suits record fixtures and individual tool responses. Parquet can carry larger relationship and event tables for offline analysis. A schema should define field types, nullable values and identifier rules so that converting between formats does not change meaning.

The environment can expose `read_record` and `read_anchor` actions, with evaluation hooks checking that submitted citations resolve. Reset fixtures should restore the same identifiers and content digests for a repeated case. If rendering changes page boundaries, stable logical anchors should remain usable. This prevents a presentation change from breaking a valid evidence reference or accidentally pointing it at different content.