I don't trust the LLM to classify my email. So I don't let it.

Short summary

Separate LLM feature-scoring from decision-making: the model reads an email and returns four numeric scores (confidence, senderTrust, reversibility, urgency) but never decides—that happens via explicit rules in a single file mapping scores to actions (PUSH, QUEUE, SILENT, AUTO). Benefits include auditability, cheaper models becoming optimal rather than a compromise, controlled blast radius, testable logic, and stable learning. When the LLM fails, keyword fallback scores work the same way.

• •Separate LLM scoring (what it perceives) from decision logic (what the system does)
• •Use explicit rules in code, not prompts, for decisions—enables auditability and testing
• •Cheaper models become optimal because they read signals consistently, not because they're a cost compromise

Generated with AI, which can make mistakes.