What technical handover makes a complaint pilot reproducible?

A pilot manifest can identify `release_id`, `schema_version`, `environment_version`, `evaluator_version` and the applicable licence reference. File entries can include content hashes and declared encodings so the lab can verify that an imported export matches the reviewed material.

For executable tasks, request a starting snapshot identifier, a reset contract and any generation or simulation seeds. A seed alone is insufficient if tool behaviour, policy files or evaluator logic changes, so those dependencies should also be pinned. Reset documentation should state whether it clears action logs, queued events and simulated clock advances.

A replay record can capture the task configuration, model actions and tool responses. If the environment includes non-deterministic behaviour, ask which parts can be replayed exactly and which require comparison within declared bounds.

For MCP delivery, a buyer can request version information alongside retrieved material so a later export can be matched to the same release. The first integration exercise should load, run, reset and rerun a selected task, then compare the relevant state and score outputs. Any differences should be attributable to documented variability or a specific version change.