AI safety and governance increasingly demand that systems' normative judgments be transparent, testable, and auditable, yet deployed alignment overwhelmingly reduces moral evaluation to a scalar---a reward score, a safety probability, a guardrail pass/fail---which both discards moral structure and silently pre-commits to a single ethical framework while appearing neutral. We present , a structure-preserving compiler that parses natural-language scenarios into a typed, canonically-hashed intermediate representation and runs a plurality of framework analyses over it (consequentialist, Kantian with an SMT-checked universalizability gate, virtue, care), refusing to aggregate when frameworks disagree; and (Democratically-Governed Ethics Modules), an evaluation engine mapping situations to a canonical nine-dimensional moral vector and per-stakeholder tensor, with constitutional modules that hold hard veto. We frame the methodology as : converting normative claims into explicit models with declared invariance assumptions, extracting measurable predictions, and testing them. We report reliability , a measured safety property (robustness to meaning-preserving manipulation, surfaced as invariance violations), and pluralist governance (a worst-off/escalate policy that routes genuine disagreement to human review). Both systems are open and installable. Target venue: AAAI/AIES. Author preprint deposited for archival and citation. Draft — pending author review.
Andrew Bond (Sun,) studied this question.