How can generated change material be packaged for an agent harness?

A dataset can use Parquet for linked state tables and JSONL for request events and task definitions. A task record could contain `task_id`, `initial_state_id`, `request_event_id` and `allowed_tool_set`. Stable identifiers let a harness assemble a task without depending on file order or display labels.

Tool contracts can define actions such as `read_agreement`, `list_open_requests` and `submit_change`, including required arguments, return types and structured errors. Field descriptions should explain business meaning as well as data type; a string labelled `status` is insufficient without its permitted values and transition rules.

A separate evaluator manifest can reference expected state constraints and hidden facts. Those artefacts should remain outside the agent's retrieval paths and tool responses, preserving the intended information boundary.

The package can also declare schema versions and referential integrity rules so ingestion failures are distinguishable from agent failures. Buyers can ask for a compact loader example that reconstructs a task, exposes only authorised observations and records the resulting tool trace for later inspection.