The BlueVeta team explained that the fluctuation in ChatGPT's accuracy between the two batches is because the consumer chat interface does not provide a fixed model snapshot guarantee, and backend routing may change between sessions.