How can data structure expose the difference between timely evidence and later reconstruction?

A generated release can distinguish `event_at`, `recorded_at`, `available_at` and `review_cutoff` rather than place every temporal fact in one timestamp. These fields answer different questions: when the action occurred, when a system captured it, when the reviewing role could retrieve it and which moment the task asks the model to assess.

Consider authorisation record AUTH-2187. Its text describes an approval before a purchase, but the record enters the system after the review cutoff. The model may use it in a retrospective review if the task permits that evidence. It cannot silently treat it as material available to the original reviewer.

JSONL records could carry these timestamps alongside `record_id`, `source_system` and `timestamp_basis`. An explicit timezone and a documented rule for date-only records prevent accidental ordering assumptions.

The task interface can accept an `as_of` argument and return only evidence available at that point. Comparing answers before and after a late record becomes available tests whether the model revises its conclusion for a defensible reason, rather than allowing future information to leak into an earlier decision.