What's the difference between concept drift and data drift?

Data drift is the input distribution changing while the correct output for a given input stays the same; concept drift is that relationship itself changing, so the same input now warrants a different answer.

The standard framing: “data drift is the change in the input data distributions; concept drift is the change in relations between model inputs and outputs,” and one can happen without the other, so a model can keep performing reliably through data drift alone. For an agent, data drift looks like a new mix of requests that still get the same right answer they always did. Concept drift looks like the rule itself changing underneath the agent: a refund policy that used to allow 30 days now allows 14, so a request the agent used to correctly approve now needs a correct denial.

Tessary’s behavior_drift classifier sits closer to the data-drift side of that line: it flags a shift in the steps an agent takes without judging whether the new steps are correct, the same limit plain data drift has on its own. Telling a harmless shift apart from a concept that quietly changed is a judgment call it hands to a person.