So I built a second experiment using the same agent, the same task, and the same models. The only thing I changed was introducing conflicting instructions across the three prompt slots.