After looking at the Gemini 2.5 iterations under Appendix: “Gemini 3.0” A/B result versus the Gemini 2.5 Pro model, I couldn't help but think:

It's like a child who's given up on their homework out of frustration. Iteration 1 is way off, 2-3 seem to be improvements, then it starts to veer wildly off-track until essentially everything is changed in iteration 10. E.g. "HERE, IS THIS WHAT YOU WANT?!"

Which led me to hypothesize that context pollution could be viewed as a defense mechanism of sorts. Pollute the context until the prompter (perturber) stops perturbing.