How should review records represent what was knowable at decision time?

A review integration should distinguish when something happened from when the reviewing team could know about it. Useful fields include `occurred_at`, `available_at`, `review_opened_at` and `decision_at`. A document may describe an earlier event but arrive after the decision.

Link evidence through `review_id` and `evidence_id`, with an explicit `evidence_cutoff_at` for each task. A retrieval tool such as `list_review_evidence` can filter on availability at that cutoff. Without this distinction, a downstream system may answer an earlier review using a later correction that makes the task artificially easy.

JSONL can preserve nested evidence references, while Parquet can hold separate review and evidence tables joined by stable identifiers. Either format needs documented timestamp semantics, including how missing availability times are handled. An absent timestamp should not silently mean that evidence was always accessible.

For resellers, these semantics belong in the accompanying schema rather than only in sales copy. Customers rebuilding the material in another platform need to retain the same evidence boundary. A useful integration check compares the records returned before and after a late document becomes available.