Definition
Plain language
A crossing test where you change one thing and the behavior moves, change a different thing and it doesn't — pinning down which factor is really responsible.
As stated in the literature
An experimental design that demonstrates two variables affect different processes: manipulating factor A changes behavior X but not Y, while manipulating factor B changes Y but not X, isolating the causal driver.
Why it matters: It matters because it lets researchers pin down which specific factor causes a behavior instead of guessing from a single change.
For example, if changing the wording of a question shifts a model's answer but changing the topic does not, and changing the topic shifts something else that wording leaves alone, you've shown the two factors act on separate things.
Heard on the show
“To tell those apart he runs a double dissociation.”Episode 229 — One Word Flips a Chatbot From Backbone to Yes-Man