Content moderation is usually cast as flat multi-label text classification. That framing discards the social structure of the problem: a platform serves many stakeholders whose values genuinely differ, and a single label can neither represent nor govern that plurality. We reframe moderation as the construction and governance of a . Each stakeholder's values are encoded as an (EM)---a weighting over ethical-concern primitives that, composed in a directed acyclic graph (DAG), the framework can enforce. A conversation is compiled---once per stakeholder ethos, via a structure-preserving moral compiler with an LLM extractor---into a rank-4 moral tensor over the axes (dimension, stakeholder, turn, action, coalition, uncertainty). Per-stakeholder verdicts are aggregated by an policy: worst-off (the most protective verdict governs) and escalate (cross-stakeholder disagreement, or any escalate-tier verdict, routes to human review rather than collapsing to a confident scalar). We evaluate the system on 240 real moral dilemmas (the AITA corpus) under two shipped stakeholder ethos. Three findings: (i) the ethos encode systematically value systems, in the direction their construction predicts (e.g.\ the harm/care ethos weights physical harm significantly more negatively; all per-dimension sign tests p<0.001; verdicts diverge on 34\ (ii) the multi-dimensional tensor ---aggregate harm does not track the human culpability label ( -0.05, n.s.), but specific normative dimensions (fairness, care) do, structure a scalar toxicity score conflates; and (iii) worst-off/escalate is consequential---relative to a single decisive stakeholder, it routes 34\ than laundering, stakeholder disagreement. We report honestly that DEME is not a culpability classifier and that escalation does not discriminate among uniformly hard dilemmas. We release the full evaluation and analysis harness. Target venue: IEEE TCSS. Author preprint deposited for archival and citation. Draft — pending author review.
Andrew Bond (Sun,) studied this question.