How can a training record connect a routing decision to its supporting evidence?

A routing training record can separate the model input from the target decision. In JSONL, an input object could contain `request_id`, `message`, `context_refs` and `policy_refs`; a target object could contain `destination_id`, `decision_code` and `evidence_refs`.

For CASE-0639, a context reference might point to the account status field in ACC-2187, while a policy reference identifies the clause assigning closed-account enquiries to a specialist team. Evidence references should resolve to stable fields or text spans, rather than depend on the order in which documents happen to appear.

The same structure can use Parquet tables for requests, context records, policies and targets. Foreign-key checks can confirm that every cited item exists in the release.

MirrorFirm could generate these linked records for supervised training or retrieval-assisted tasks. An evaluation hook can inspect whether the chosen evidence supports the destination, separately from whether the destination matches the target.

Keep reference answers outside the input package. Otherwise, a retrieval step may expose the intended decision before the model has interpreted the business evidence.