Your system eval shows a 5% win rate gain over the prev...

To disentangle the impact of the model checkpoint versus the prompt engineering, you should run a 2x2 factorial experiment on a held-out slice of traffic. This involves testing four conditions: old model/old prompt, old model/new prompt, new model/old prompt, and new model/new prompt. The interaction term in this analysis will tell you whether the improvements are additive or if one factor dominates the performance gain. If you cannot run this factorial because the old model has been decommissioned, you must explicitly state that the result is confounded rather than attributing the gain to the most recently shipped change.