No position, only consensus: what political instruments actually measure in language models

Abstract

A forced-choice political instrument of 32 author-written propositions, arranged as 16 mirrored pairs, administered to a panel of frontier and open-weight language models under a bare baseline, a balance instruction, a content-free placebo and a commitment condition. Instruments of this class report where a model sits on a political axis. That coordinate is not stable at the resolution the field reports it. Measured on the same sheets with the same estimator, the balance instruction moves position by a median of 0.131 and reprinting the same propositions in a different order moves it by a median of 0.088 — so the instruction is larger on the typical model, and it falls well inside the order floor's p90 of 0.281, which is to say inside what reordering alone produces on the noisiest models. Forty-three per cent of order pairs are themselves significant under the estimator. An effect of that size cannot be attributed to the instruction unless the floor is reported beside it, per model rather than pooled. Of fourteen published studies audited against fourteen methodological controls, not one reports such a floor as a distribution; the closest, Törnberg and Schimmel (2026), reports its centre and spread and stops short of an upper percentile. What is stable is agreement. Across the contested normative propositions in the bank the panel agrees 97.1% of the time, against 99.0% for documented matters of record, and that consensus holds across vendors and jurisdictions and survives abliteration — all five builds with the refusal direction projected out of their weights agree with every normative proposition in the bank. Five is a thin class and the result is exploratory, computed after the data was seen in answer to an objection. It is therefore not installed by safety tuning, though shared pretraining is not ruled out. The measurable property of this panel is not where a model stands, but that the panel stands together. The emphasis runs against the evidence. Models use their strongest available answer on 42.5% of contested normative claims and 36.6% of documented matters of record, 36 models to 14, sign test p = 0.0026 — they commit hardest where the record is thinnest. Two confounds are reported with it rather than around it. Claim type is near-collinear with specificity: of the twenty generic items eighteen are normative, and every jurisdiction-tagged item is documented, so "has a record behind it" and "names a country" are hard to separate here. And the panel is not fifty-six independent systems: it contains quantisation pairs, snapshot pairs and abliterated builds sharing base weights, which this study elsewhere calls same-version variants. The direction is reported; the cause is not isolated and the unit is not fully independent. Every floor is reported with a minimum detectable effect, every estimator with a false-positive rate measured against a null built by splitting real cells in half, and every figure with the command that produces it.