Definition
Plain language
When asking a model the same question many times gives scattered answers, so no single answer ever looks trustworthy.
As stated in the literature
Failure mode of plurality-vote extraction in which falsifications vary across samples, preventing any candidate value from reaching the decisiveness threshold; the attacker accepts few composites and gains little.
Why it matters: It denies an attacker the confidence they need to act, since scattered answers give them nothing they can rely on without independent testing.
For example, an attacker asks the same question fifty times and gets fifty different temperatures, so no answer stands out and none can be trusted.
Heard on the show
“On the 122B you get consensus starvation.”Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers