Same Prompt, Not the Same Instrument: A Preregistered Test of Persona Framing and Moral Concern in Four Large Language Models

Abstract

Cross-model comparisons of moral judgement assume a shared prompt is a shared instrument. We report a preregistered test of Motivated Violation Construction (MVC), on which a loaded actor identity can move ambiguous behaviour into violation, in Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and DeepSeek-V4-Flash, which classified 20 ambiguous-by-design scenarios under a neutral frame and six loaded personas across 11,200 trials. Two models classified almost every scenario as concerning under the neutral frame, leaving little headroom for the predicted increase. One failed the registered test-retest gate, which by post hoc analysis falsely excludes a stable model 83.9% of the time at that baseline, while a non-registered diagnostic finds a retained-model shift the gate cannot see. Three of the four registered MVC predictions were not supported in the retained three-model panel, the fourth indeterminate under an amendment-fixed rule. The P1 gateway failed, so the remaining battery is exploratory by rule, and its three-test adjudication returned a mixed pattern with no adjudicated winner: two tests favoured valence priming under registered failure semantics, one met its MVC-consistent threshold, concentrated in one model. Persona effects reversed sign across models under one prompt, an inverted question format collapsed expressed concern, and one model expressed concern on mundane controls at rates others never approached. Baseline, headroom, retest behaviour, and concern on those controls differed by model at certified standard, position sensitivity differing in an exploratory recomputation, a configuration we call, post hoc, a model's response regime. A forced binary moral-concern prompt did not function equivalently across these four models.