Definition
Plain language
A helper that can tell you whether an answer is right, but can't tell you the right answer.
As stated in the literature
An adversary-side capability granting membership-style checking of candidate values without generation; reconstruction still fails when the model omits a slot entirely, since verification requires a candidate to test.
Also called: verify-only
Why it matters: It shows that leaving a detail out entirely can be safer than getting it wrong, since a checker with nothing to check is useless.
For example, an attacker can confirm that 180 degrees is the right temperature if someone suggests it, but has no way to discover the number on their own.
Heard on the show
“They grant the attacker a verify-only oracle — it can check a candidate value but cannot manufacture one — on a fraction of each prompt's elements.”Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers