Definition
Plain language
A measure of how much extra, tangled-up activity is going on inside a model beyond its main line of thought.
As stated in the literature
In the Rift deception work, a metric for how much competing structure remains in a model's internal state once the dominant directions are accounted for; deliberate deception raises it about two-fold over an identically-wrong honest answer.
Why it matters: It offers a measurable internal signal that can distinguish deliberate deception from an honest error that looks identical on the surface.
For example, when a model knowingly lies, more competing activity churns underneath its answer than when it is simply mistaken.
Heard on the show
“The metric — they call it residual rank — is basically: how much is going on beyond the lead melody?”Episode 153 — Catching a Lie From the Inside, When the Words Look Completely Honest