Why should routing data preserve event order instead of only final queue labels?

A final queue column hides the sequence that produced an assignment. A request may first require clarification, then become eligible for specialist handling after a response arrives. Training only on the final destination can teach an agent to skip the necessary question.

A generated release can represent this sequence through linked event records. Useful fields include `event_id`, `request_id`, `event_type`, `occurred_at`, `visible_at`, `actor_role` and `payload_ref`. The distinction between occurrence and visibility matters when an upstream system reports an event late.

JSONL can preserve ordered interactions, while Parquet can support analysis across event types and destinations. A request table can hold stable identifiers; an event table can hold the changing evidence. Explicit foreign keys should connect both to the relevant policy version.

For training, construct observations from events visible before the decision point. For evaluation, compare the agent's action with the permitted transitions at that point, rather than the eventual case outcome.

This representation supports a more precise question: was the route justified by available evidence when the agent chose it?