Small local models: what actually holds up when you re-run the measurement

I run a small local model — Qwen3.5 4B via Ollama — as one deterministic-checked step inside some...