LLM judge

Formal verification by agents is harder than it sounds

The author compares agent-based formal verification to asking an LLM judge to fix LLM output, arguing interpretable, expressive, extensible formalisms remain extremely hard.