Definition
A small detector that reads a model's internal state to spot when it's recalling something that's gone out of date.
A linear classifier over a model's hidden states trained to predict whether a fact has been invalidated by world changes after training cutoff, used as a deployment-time staleness signal.