A language model on its own can write a sentence, but it cannot guarantee that the sentence contains the exact form you asked for. So inflection is not the model's job at all: the forms come from rules, the model only writes the sentence.
Between the model output and the response there are gates. They reject a sentence mechanically if the target form does not appear in it, if length or register is off, or if the structure is impossible.
Every prompt change is run as an A/B series: the decision rule is written down before the run and the result compared against a hand-labelled anchor set, not against an impression of how the samples looked. Several promising-sounding ideas have died in that measurement and been rolled back. An empty response beats a wrong sentence.