A small, throwaway build whose only purpose is to answer one technical question: can this integration work, will this model be accurate enough, will this load be handled. It is not a product and is not meant to ship.
Some risks cannot be resolved by discussion. A week spent proving the hard part works is far cheaper than discovering in month four that the core assumption was wrong.
We define the exact question, the success criterion, and a time box before starting. The result is a short report with the answer and a recommendation — and we say plainly when the answer is no.