Is this some kind of calibration then? I'd expect that the probabilities automatically adjust during training, such that in "lock" mode, for example, syntax-breaking tokens have a very low probability and would not be picked even wich higher temperature.