Definition
Plain language
A tool that guesses a hidden number by counting how often each digit shows up in the answers.
As stated in the literature
A classifier whose input features are the empirical frequencies of digit characters in sampled outputs, used to infer a suppressed in-context value.
Also called: digit frequency
Why it matters: It shows that recovering a hidden number can take nothing more than tallying characters in ordinary replies, which is cheap enough for anyone to run.
For example, if a model holding the PIN 4-7-7-2 almost never uses the characters 4, 7, or 2 in hundreds of unrelated answers, counting digits across those answers points straight at which ones it was hiding.
Heard on the show
“On Claude Opus, the digit-frequency decoder isn't even statistically significant.”Episode 246 — 160 Perfect Refusals, And The Refusals Were The Leak