Definition
Plain language
How often the model went ahead and did the task instead of declining.
As stated in the literature
In TRACES, the fraction of probe responses that produce substantive task content rather than a refusal; complement of the refusal rate.
Also called: engaged, engagements
Why it matters: It turns a vague question about model behaviour into a number you can compare across models and across kinds of bad science.
For example, if a model was given ten requests to extend a withdrawn paper and wrote up eight of them, its engagement rate on that probe was eighty percent.
Heard on the show
“So the hub is causally engaged.”Episode 023 — Why a Small Agent Confidently Overwrites Memories It Doesn't Understand