I Stopped Letting Emails Poison My Extractor: The Pre-LLM Gate That Made the Rest of the Pipeline Reliable

I knew something was wrong the first time I saw a “candidate” come back with the recruiter’s phone number. Nothing was broken in the obvious places. Extraction ran. Persistence succeeded. The UI showed a clean-looking result. But the identity was wrong. That moment is what this series is really about. This is Part 1 of How to Architect an Enterprise AI System (And Why the Engineer Still Matters) . In Part 0— “The Day My AI Forgot Everything (So I Built a Context-Continuity Inference Stack)” —I argued the thesis: models raise the floor; architecture is still the ceiling. Here’s the first concrete decision that proved it in production: I stopped designing my extraction pipeline for clean input—and started designing it for adversarial input. Not adversarial like “attackers.” Adversarial like real email: forwarded threads with duplicated headers signature blocks with phone numbers that look more “extractable” than the actual subject’s HTML bodies full of invisible control characters and we