#Measurement

We measured how frontier LLMs frame politically contested institutions, escalating from a prompt edit to an elicitation pipeline to cutting the refusal reflex out of the weights. A 2026-08-31 re-measurement on a forced-choice instrument narrows the central claim: the politeness layer suppresses whether a model will commit, not which side it lands on. It also establishes the noise floors this kind of measurement has to clear — presentation order alone moves more answers than any manipulation we tried. Full data, scripts, protocol public at github.com/gorrie/bias-study.