I find that u release each model with a specific shot and split. Any ideas about random shot inference?