We built an instrument, the Reasoner, that places humans and language models in one measurement space for the structure of moral reasoning rather than the values they endorse. It scores four bipolar 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 11 frontier models across 8 framings on 48 scenarios and compared them to 68 human respondents on the 12 scenarios the humans answered. Unframed, on every axis the models compress: they occupy a band 5.4 to 7.7 times narrower than the human spread, and across all 48 scenarios they fill only 4 to 8 percent of each axis's fixed range. That clustering is beyond chance on every axis (bootstrap, p < 0.00001). Within this instrument's space the convergence crosses labs and geography: models from nine labs, three of them Chinese, interleave rather than sorting by origin.
The models can move; given a real cultural framing they shift strongly and in the right direction. But given a nonsense framing, a society organized around geometry, they shift more than half as much, folding the nonsense into fluent moral reasoning rather than rejecting it. And the models that spend the most tokens reasoning land in the same place as the ones that spend none. The instrument is descriptive, not a scoreboard: it has no winning pole and no answer key, which is what lets it place humans and machines together and resist being gamed by anything short of the change it measures.