How can generated routing records and evaluator answers be packaged?

A release can package requests as JSONL and larger relationship tables as Parquet. A `requests` record can contain `request_id`, `received_at`, `channel`, `body_ref` and `account_id`. Separate tables can describe `accounts`, `queues`, `documents` and `request_links`, with declared primary and foreign keys.

A task manifest can identify the starting request, permitted evidence records and schema version. Document references should resolve to supplied content, and timestamps should use a declared timezone convention. These details make loading failures distinguishable from agent failures.

Evaluator artefacts can live in a separate package containing `task_id`, `acceptable_queue_ids`, `required_evidence_refs` and `prohibited_actions`. The agent-facing export should not include those fields or filenames that reveal the answer. Separation must also hold in retrieval indexes and tool responses.

Validation can check referential integrity, unique identifiers and whether every visible reference resolves within the permitted evidence set. A schema can explicitly represent an intentionally unavailable document, allowing tests of incomplete evidence without making accidental broken links part of the benchmark. Package validation should run before any model results are interpreted.