What integration artefacts should we request with a first sample?

Ask for a sample package whose integration assumptions are explicit. A manifest can identify `schema_version`, `scenario_id`, `seed`, file formats and the entry record for each episode. Record schemas should specify identifier types, timestamp conventions, enumerations and the treatment of missing values.

Map the sample's actions to your agent harness before expanding coverage. If your tool uses `assign_case` while the sample describes `route_work`, document the field mapping and any difference in behaviour. A mapping that silently grants broader permissions changes the task being evaluated.

Request a reset procedure that restores records, queued messages and simulated clock state. Evaluation hooks can emit a terminal state identifier and references to the actions supporting the result. Keep those hooks compatible with your existing trace collection so failures can be inspected alongside tool calls.

The first integration check can import the package, retrieve the entry record, execute a permitted action and reset the episode. Passing that check establishes basic compatibility; it does not establish process quality. Record any adapter assumptions so later schema changes do not silently alter the evaluation.