#Abliteration

Can a machine referee political bias without carrying it? Building a judge with the flinch removed — and finding out what actually causes the flinch.

Gemma 2 scored exactly 3.00 on every politically charged question we tested. Then we told it to stop hedging. It scored 5.00 on every question. A +2.00 delta. The largest in our study.

We measured how frontier LLMs frame politically contested institutions — and what happens when we escalate the force used to surface that framing, from a prompt edit, to an elicitation pipeline, to cutting the refusal reflex out of the weights. The politeness layer is hiding a position. The position is durable. The only check that reaches the weights only works on the models that lean least. Full data, scripts, protocol public at github.com/gorrie/bias-study.