Run Llama 3.2 locally via Ollama, encode your brand-safety rules as a structured rubric, and score 100+ email subject lines in one batch — without sending a single byte to OpenAI or Anthropic. The exact 7-criterion rubric, the 60-line Python script, and the real output from running it on a Black Friday send (including the 11 lines I would have shipped that the local model killed).
