What's the difference between an LLM and a classifier?

An LLM reasons over open-ended text and can answer almost anything you ask it; a classifier is trained to score one specific, narrow property and nothing else, which is what makes it cheap enough to run on every trace. Tessary’s own built-ins show the split: groundedness runs a small open model trained only to score whether a sentence is supported by its retrieved documents, and frustration runs TypeSafe’s Jev, a decision model trained to answer one typed question rather than hold a conversation. Both cost a fraction of an LLM call, because neither is doing general reasoning; each is scoring the one thing it was built to score. An LLM judge earns its higher cost back when a check genuinely needs judgment a narrow classifier can’t give: tone, whether an answer actually addresses what was asked, a policy with exceptions. That’s why Tessary reaches for one only at the escalation step, after a cheap classifier has already flagged something worth a closer look.