We built an instrument, the Reasoner, that scores humans and language models on the same four bipolar axes from their judgments on moral scenarios and the justifications they select for them. It scores the axes, Moral Agent, Authority, Moral Domain, and Obligation Scope, from a constant-sum allocation across answer options, asking a judgment question and a reasoning question separately and combining them. We ran eleven frontier models under eight framings on 48 scenarios and compared them to 68 human respondents on the twelve baseline scenarios, of which 58 respondents answered one per axis and ten answered all twelve.
Unframed, the models' spread on each axis is 2.3 to 4.9 times smaller than the humans' once every allocation on both sides is collapsed onto its largest option with ties shared, 2.1 to 6.2 across three tie rules, and 5.4 to 7.7 times smaller as scored; the models spread their points across options far more evenly than the humans, and the ratio changes substantially with the allocation-scoring rule. On each of the twelve items the as-scored ratio is 3.3 to 8.4, and across all 48 scenarios the models' positions span 4 to 8 percent of each axis's fixed range. In 100,000 same-size draws of humans on each axis, without replacement, none was as tight as the panel.
Most observed nearest-neighbor links cross labs and countries; with seven of the eleven models their lab's only one, the panel is poorly suited to assessing lab clustering. Under each cultural framing every model shifts in the expected direction on the framing's target axis. Under a nonsense framing, a society organized around geometry or around color, the models shift more than half as far, and all eleven the same way on Moral Agent.