A Framework for Pluralistic Value Assessment (MindOrder)

Abstract

We present a computational framework for multi‑dimensional value assessment, designed to address a structural problem shared by historical evaluation and AI value alignment: how to adjudicate among incommensurable values without collapsing them into a single score. Mainstream approaches to value aggregation seek a correct value function — a utility function, a preference model, or a scalarized objective. We argue this framing is structurally inadequate. We propose instead an architecture of two layers with fundamentally different epistemic status: 1. Contestable standpoints (relative). Value inputs — cultural perspective, historical standpoint, attribution weights — are modeled explicitly as declared standpoints. Different evaluators may hold different standpoints and reach different, mutually incompatible conclusions. This is a feature, not a defect: the framework's output is not a single verdict but a set of traceable judgments, each valid within its stated standpoint. 2. Invariant constraints (absolute). A small set of axioms — chiefly the inviolability of noncombatant life — acts as a hard constraint on the inference procedure. Standpoints may not override it; probabilistic aggregation may not trade against it. We formalize this as a five‑stage adjudication procedure over four independent dimensions (Ci / Zhi / Bo / Gong — roughly: motive, efficacy, cognition, execution), implemented as a runnable engine and evaluated on 172 historical subjects spanning seven subject types (persons, institutions, intellectual movements, laws, legal families, states, organizations). We report three classes of validation: (i) internal consistency, (ii) parameter sensitivity, and (iii) ablation. We state explicitly that internal consistency (172/172; 6/6 for edge cases) demonstrates coherence, not validity; external validity requires independent raters and is left to future work. The ablation pass drove a genuine self‑correction cycle: it detected that several headline rules were not carrying load as implemented, traced the root cause to suspended dimension levels that never reached the output, and repaired both defects (dimension‑level surfacing; C8 name–reality alignment). We close by stating what evidence would falsify the framework. Keywords: value pluralism; AI alignment; computational ethics; historical evaluation; declared standpoints; non‑aggregation The implementation, case file, and interactive interface are available in the companion GitHub repository. Generative‑AI disclosure: generative language model was used for language polishing and structural suggestion; all substantive content is the author’s own work. This is draft v3.9 technical report prepared for preprint submission.