# What data trains a custom classifier?

Your own agent's production traffic, not a generic public dataset. That's the point of building one instead of relying on a built-in: a public dataset teaches a check what refusals or errors look like in general, but your agent's actual failure modes, the policy it has to follow, the tool call it keeps getting wrong, only show up in your own traces.

The examples that shape it come from the behavior you want caught, in your own traffic, so what it scores is the failure you actually have rather than a generic proxy for it. Your traces stay yours either way: Tessary doesn't use them to train anything shared across customers.

---

Source: https://tessary.ai/answers/tessary-custom-classifiers/what-data-trains-a-custom-classifier
More on Tessary custom classifiers: https://tessary.ai/answers/tessary-custom-classifiers
From Tessary, agent reliability for AI agents in production: https://tessary.ai