See the private-data handoff before building your own.
Change the threshold, run 200 generated event rows, inspect the exact local outcome, then choose the coarse fields that may leave the page. This sample uses no customer data and makes no network-isolation claim; the versioned Runner enforces that boundary for real evaluator packs.
Your fit question next
Define the contract, rehearse it locally, then decide whether the reusable pack is worth exporting.