The architectural design principle of a frozen/mutable boundary in Dream-RSI.

First, thanks for the paper and the release plan — the harness-level framing (evolving the exploration policy around a frozen base model rather than the weights themselves) is a clean, legible design, and it's a genuinely useful reference point for thinking about where self-improvement loops should and shouldn't touch the underlying model.
I'm working through a few recent self-evolving-agent papers (Dream-RSI plus the four you cite in §6 — Meta^n, Darwin Gödel Machine, G-Zero, R-Zero) and comparing how explicit each one is about the boundary between what stays frozen and what gets optimized across iterations. I wanted to raise a question about Dream-RSI specifically, phrased as a request for clarification rather than a claim that anything is wrong with the design.
As I read it, the paper is very clear and consistent about the boundary itself: the base model (Gemini-3.1-Pro / Gemini-3.7-Flash), the evaluator, and the execution interface stay frozen; only the exploration-policy code is what actually gets optimized round over round. That's stated repeatedly and unambiguously — no confusion there.
What I couldn't find, and would be glad to be pointed to if I missed it, is any place where the paper (or the repo) states why that particular split is the safe one to hold — i.e., a Limitations section, a Broader Impact / safety-considerations section, or even a paragraph arguing that keeping the evaluator and base weights frozen is what prevents the loop from drifting or reward-hackin...