How could generated handover records be packaged for a benchmark pipeline?

A generated release could use Parquet tables for cases, transfers, dependencies and attachments, alongside JSONL task manifests. Each table should declare its primary key and foreign keys. A transfer record might contain `transfer_id`, `case_id`, `sender_team_id`, `recipient_team_id` and `context_bundle_id`.

Represent dependencies as explicit edges with `prerequisite_id`, `dependent_id` and `required_state`. Avoid making the loader infer a blocking relationship from a phrase in a note. Attachments can retain readable content while exposing document identifiers and links to the records they support.

Each task manifest could specify `task_id`, `initial_snapshot_id`, `entry_record_id`, `allowed_tools` and an evaluator reference. Keep expected outcomes in a separate evaluator package so they are not accidentally included in the agent's retrieval index.

A schema manifest should identify field types, null semantics and enumerated states. Loader checks can reject broken foreign keys or duplicate identifiers before a benchmark run begins. This packaging makes structural faults distinguishable from agent errors and lets buyers inspect the same case through both table queries and rendered business records.