Position Bias Masquerades as Moral Steering in Multilingual Forced-Choice Evaluation

Prior work steers moral frameworks across languages and reports that lower-resource languages are harder to steer. We ask whether that reading survives controls for instruction following, answer validity, and option-order sensitivity. On the MultiTP trolley-problem corpus we run seven model-mode cells, each with 19,320 generated responses across six languages and seven conditions. Under the standard forced-letter format, baseline same-pole agreement under a label swap falls below the 0.5 reference in six of seven cells, and any reported consistency rate decomposes into complementary same-letter and same-pole readings. In a baseline-only ablation on two models, requesting the chosen option text instead of its letter raises paired same-pole agreement by $+0.27$ to $+0.62$ and recovers parseable Zulu pairs from a model that produced none under forced letters. Restricting to order-stable pairs retains 5.5 to 43.0% of scheduled pairs, moves resource-gap estimates in both directions, and leaves interval-exclusion counts that depend on the resampling unit (8 of 28 with independent records, 11 of 28 with dilemma clusters), which we report as exploratory. A pre-specified safety comparison on one cell gives a Zulu-minus-English attack-success gap of $+0.083$, below its 0.15 threshold. Multilingual steering scores should be reported with response validity, semantic order sensitivity, selection coverage, and the scope of their controls.