Would constraining the tokens that the model can choose from impact its “intelligence”/“creativity” in other (possibly negativel) ways?