The variance test that flipped my local LLM ranking

I've been picking models for a small Ollama pool that handles delegated coding chores from my main agent. Parsers, recursive transformers, throwaway test scaffolds — the kind of thing where round-trip