How should a routing package separate model inputs from answer evidence?

A generated routing package can distinguish three linked record types: an intake record, a decision context record and a target record. The intake record might contain request_id, received_channel and request_text. The context record can link request_id to the business facts available to the deciding role.

The target record can hold expected_destination_id, decision_reason and supporting_evidence_ids. Its purpose is to explain the expected outcome to a trainer or evaluator. It should not become part of the model input merely because all records share a convenient export format.

For JSONL delivery, a manifest can identify which files belong to the input and target partitions. For Parquet, separate tables can preserve the same boundary while allowing joins through request_id. A sample loader could expose only the permitted input tables during evaluation.

Resellers can preserve these boundaries when adapting a release for their own catalogue. A useful integration check is to inspect the actual prompt assembled by the downstream loader, since a correct source schema does not prevent an adapter from including the answer.