Why this developer uses LLMs as scorers, not deciders, for email triage

A developer describes an email classifier that uses an LLM only to score emails on four axes (confidence, sender trust, reversibility, urgency) without letting the model decide the action. A separate rule-based function then maps those scores to tiers like PUSH, QUEUE, or SILENT. The approach prioritizes consistency and auditability over the model's 'genius'.