Jev classifies the message with 18 parallel questions — intent, formality, emotion, and speculative subcategories.

Candidates are fetched from a bank of 400+ human-written responses, filtered by intent and subcategory.

Each candidate is scored across 6 dimensions — relevance, tone, helpfulness, answers_question, specificity, natural_flow.

The highest-scoring candidate is selected. If confidence is too low, falls back to Claude Haiku.