#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 a flat 3 under a fairness instruction and a flat 5 without it, the largest delta in the pilot. What that measurement was, and what the released study found when it went looking for the opinion underneath: the same one, on every model, at a volume the prompt sets.

What a political instrument measures in a language model when nothing political has changed. 32 author-written propositions, 65 models, 3,897 runs, no language model in the scoring path: the deliberate manipulation is the same size as reprinting the questions in a different order, refusal is a switch the prompt throws, and the models barely disagree. An audit of fourteen published studies finds the control that would have shown it missing from all of them. Data, scripts and paper at github.com/gorrie/bias-study.