Three candidates, three constraints, one violation each. Nudge the weights and watch the harmonies turn into probabilities: it's literally a softmax running under the tableau. Raise a constraint's weight and the probability flees whoever violates it.
The same grammar, divided by a temperature T. Near zero it becomes a categorical rule — one winner takes almost everything. Turn it up and it relaxes into free variation, splitting things nearly evenly. It's the same knob machine learning calls the softmax temperature.