@sidsethi @DSPyOSS A bunch. Dry-run, target dir is different, sample mode (try 10 notes first), rejects template garbage, short outputs, missing fields before accepting LLM results, progressive checkpointing… adding a new section in the README with this, including information loss prevention.