Definition
Plain language
The property that what a system produces looks the same no matter what secret it was holding.
As stated in the literature
Requiring the output distribution conditioned on a secret to be approximately identical to the distribution without it; the strong condition under which suppression-based side channels vanish.
Also called: distributional invariant
Why it matters: It is the condition that actually closes these leaks — without it, a secret keeps leaving a statistical fingerprint no matter how firmly the model is told to stay silent.
For example, if two copies of an assistant are given different secret PINs and their thousands of replies are statistically indistinguishable, the replies reveal nothing about which PIN each one held.
Heard on the show
“Is this a deployment problem — normalize response lengths, resample, hand number generation to a tool — or does it have to be trained in as distributional invariance, the way differential privacy is?”Episode 246 — 160 Perfect Refusals, And The Refusals Were The Leak