Definition
Plain language
A small detector that reads a model's internal state to spot when it's recalling something that's gone out of date.
As stated in the literature
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.
Why it matters: It offers a way to catch stale knowledge before it reaches the user, without needing to re-verify every fact through external retrieval.
For example, a small classifier reads the model's hidden state for 'who is the CEO of Acme Corp?' and flags that its stored answer is probably out of date.
Heard on the show
“The drift probe hits between eighty-three and ninety-five percent accuracy.”Episode 037 — Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say