When the same image is annotated twice, with $a$ and $b$ instances, a prediction with $m$ instances can be scored against both. Its true positives against the first annotator are at most $a$, and against the second at most $m$. If the annotators disagree in their counts, $a < b$, the total number of true positives stays strictly below the combined denominator $m + (a + b)/2$, by a margin of $(b − a)/2$: half the disagreement.

Since each matched IoU is at most $1$, the combined numerator is at most the number of true positives, so the combined Panoptic Quality is strictly below $1$ for every prediction. The ceiling is set by the annotators, not by the model.