The Moral Spectrum Analyzer: A Pre-Registered Instrument for Measuring, Validating, and Discovering Moral Dimensions in Language

Systems that moderate content or evaluate agents increasingly reason about *moral* properties of text — harm, fairness, autonomy, honesty — yet the scores they emit are typically **asserted rather than measured**: a classifier is trained on one corpus, reports a high in-domain AUROC, and is deployed as if it had captured a construct. We show that this practice is **unsound as measurement**, and we offer an instrument that replaces assertion with measurement. We define the **MoralVector**, a per-item representation with one validated channel per moral dimension — nine canonical dimensions on a frozen 3×3 ontology plus a small number of validated *extension channels* — each cell carrying a signed valence, a confidence, and machine-checkable validation provenance. We then define the **Moral Spectrum Analyzer**, the measurement apparatus that produces and audits the vector: (i) a **pre-registered cross-dataset gate** in which a dimension is "validated" only if a held-out encoder beats *both* an untrained-encoder null and a bag-of-words null by a fixed margin, with all nulls frozen before training; (ii) **joint cross-corpus training** that removes the single-corpus surface shortcut which we show inflates in-domain numbers by up to 0.44 AUROC (autonomy) — but as little as 0.01 (privacy), a heterogeneity that is itself informative; (iii) an **invariance layer** with a measured decision-stability criterion; (iv) a **discovery loop** that surfaces missing dimensions and an **admission criterion with teeth** — over the program's history it has validated one discovered dimension, retracted another, and declined a third as non-moral policy. Our central empirical result is a **dissociation between moral valence and moral identity**. Using a domain-adversarial ablation and an independent geometric ablation, we find that a dimension's *valence* (is it upheld or violated?) is register-invariant — it transfers across corpora with or without adversarial de-confounding — whereas a dimension's *identity* (which foundation is engaged?) is, for several foundations, register-bound: it occupies the *same* angular subspace as corpus/register and cannot be separated from it by adversarial training or by principal-component removal. Three complementary probes converge on this conclusion — two share a training pipeline and differ only in a domain adversary, the third is a fully independent geometric ablation on frozen embeddings. The finding has a direct engineering consequence — presence/identity channels must be built with a specific (non-adversarial) configuration — and a methodological one: **cross-dataset transfer, not in-domain accuracy, is the minimal evidence that a moral-dimension probe measures a construct at all.** --- Target venue: IEEE. Author preprint deposited for archival and citation. Draft — pending author review.