Model comparisons get noisy when prompts differ. I trust the run that keeps input, tool limits, and success criteria visible.