Anticipatory token selection was obvious when you used structured output on gpt 3.5.

The model would construct/hallucinate pre-conditions to satisfy some final output value that the model was predisposed to.