jev pipeline

I have scanned the ./src/classifiers directory. I identified three LLM calls used for classification and evaluated their fitness for Jev.

Pipeline Analysis Report: ./src/classifiers

Drafted Noul Question (spam_filter.py)

{
  "question_text": "Does this message contain unsolicited commercial content or deceptive phishing attempts?",
  "type": "noul",
  "state_keys": ["message_body"],
  "rules": ["Ignore internal routing headers", "Focus on user-facing intent"]
}

Drafted Choice Question (support_triage.py)

{
  "question_text": "Which department is best suited to handle this customer inquiry?",
  "type": "choice",
  "options": ["billing", "technical", "sales", "account_recovery", "other"],
  "state_keys": ["ticket_subject", "ticket_body"]
}

Next steps

1. Run python -m jev_pipeline lint on the generated drafts in ./jev-pipeline-out/ to verify escape options.
2. Deploy these questions to a staging environment to collect RESPONSES.jsonl for threshold calibration.
3. Use python -m jev_pipeline spread once you have 100+ logs to check for constant-output bias.