My LLM Critic Flip-Flops on Every Run. That's Fine — Because a Frozenset Decides What's Fatal.

This article explores a design principle extracted from fixing a flaw in an LLM critic. The critic returns different verdicts and explanations every time it is run on the same input. The article explains that this non-determinism is due to the architecture assigning different responsibilities to different components.

The "dangerous" direction is under-claim, where a defective plan gets zero blockers. This is the critical path where safety is paramount. The "noisy" direction is over-claim, where a sound plan gets an unnecessary blocker. These are handled by different authorities in the system - deterministic gates enforce under-claim, and a code-enforced allowlist enforces over-claim.

This allows the LLM to be non-deterministic in its severity judgments, while the code ensures the verdict aligns with structural properties. The researcher calls this the "frozenset" mechanism - the LLM proposes a severity, but code checks whether the structural property supports it. This way, the LLM's label is decorative while the family it belongs to is load-bearing.

The key takeaway is that when LLM judgment is placed on a critical path, it inherits the vulnerabilities of LLM judgment, but separating the critical path into deterministic code and advisory LLM judgment ensures consistency and safety.