Why does the timing of visible evidence matter in training records?

A business event can occur before its record becomes visible to an assistant. A customer response may reach a shared mailbox before a case system imports it. Training on the final case record can accidentally give the model knowledge that was unavailable when follow-up was due.

A generated release can distinguish `occurred_at`, `recorded_at` and `available_to_agent_at`. An action record can include `decision_at` and `observation_snapshot_id`, linking the decision to the exact evidence exposed at that point.

JSONL can hold individual decision examples with their visible observations. Parquet can hold an event table for reconstructing case history. Both representations should preserve identifiers so a buyer can trace each observation back to the generated event that supplied it.

An evaluation hook can reject rationales that cite records arriving after the decision boundary. This makes the task more faithful to operational uncertainty: the model must choose a defensible action using current evidence, rather than benefit from a completed history assembled after the outcome is known.