Noise floor (replicate variance)
The pooled variance among repeated runs of the same judicial posture on the same question, used as a measurement instrument's estimate of its own unreliability and subtracted before any openness is reported.
First used in Ariadne's Thread
Language models give different answers to the same legal question even at deterministic settings, so a posture's score is a noisy draw. Averaging several replicates shrinks that noise but does not remove it: the naive variance of posture means overstates the systematic posture effect by the noise divided by the number of replicates. Subtracting exactly that amount gives an unbiased estimate.
The floor must be estimated from replicates of the panel postures themselves. In the paper's smoke test, the panel postures ran once each and the floor came from two separate control profiles, a Median Judge and a mood-perturbed Before-Coffee Judge. The authors therefore subtracted the whole floor rather than a fraction of it, and disclosed the deviation from their own protocol.