How should a pilot package support installation, reset and later comparison?

Ask for a package manifest containing `release_id`, `schema_version`, `tool_contract_version`, `policy_bundle_digest`, `initial_state_digest` and `evaluator_version`. These fields identify the components used in a run without relying on a folder name or download date.

The installation contract should state how records are loaded and references validated before a session starts. JSONL event fixtures and Parquet state tables can coexist if the manifest identifies their roles and loading order. Reject incompatible schemas before creating partial environment state.

Define `reset` as restoration of a named initial snapshot plus any documented seeded variation. A reset response can return `session_id`, `snapshot_id` and `seed`, allowing an evaluation record to identify its exact starting conditions.

Keep a package change log that distinguishes content corrections from changes to business rules or scoring. Results from different evaluator versions may require rerunning before comparison. A useful first integration step is to load the sample, execute its supplied walkthrough and reset it, confirming that the restored state matches the declared snapshot.