Why should referral data preserve event order rather than only final account status?

A final account snapshot can conceal the reason a referral was appropriate. An account may end with specialist support accepted, but the agent initially saw only a customer message and an unresolved payment enquiry. Training against the final state can accidentally give the model knowledge it would not have held at the decision point.

A generated event stream could include event_id, case_id, event_type, occurred_at, recorded_at and actor_role. Separate customer_statement, referral_requested and specialist_accepted events make the sequence inspectable. The distinction between occurrence and recording time also supports cases where a telephone note is entered after another team has acted.

JSONL can preserve events as individual records, with case_id linking them to account and policy tables. A task constructor can expose only events recorded before the decision cutoff, while an evaluator retains the later outcome.

For resellers, ordered delivery adds value beyond a flat status column. Customers can train on progressive disclosure, replay a case at different points and check whether a model relies on evidence that was actually available. Catalogue metadata should explain the cutoff convention so customers do not unknowingly introduce future information into their evaluations.