Comment by jeffharris

thanks for flagging ... number fidelity (especially on languages that are unfortunately less represented in training data) is still something we're working to improve