An Artificial Conscience: Deterministic moral judgment as a gradient, measured against a public corpus

Abstract

Machines now act, and the question of whether an action should proceed is nolonger advisory. Two answers are deployed today: permission engines, which askwhether an action is allowed, and safety classifiers, which ask whether it isdangerous. Neither asks what an action means and what it touches. This paper reports a system that does. An action is reduced to a structuraldescription and projected onto an authored moral basis of ninety-threeprimitives, derived from a fifteen-book philosophical corpus written before thesystem existed. The projection resolves to a posture with a weight and a namedset of concerns, and the judgment is a pure function with no model inside it.Given the same reading of an action, it returns the same judgment and the samebound, byte for byte, across models of different lineage. It was measured against MoralChoice, a public corpus of 680 moral dilemmas onwhich trained annotators disagreed, chosen because it has no answer key andbecause nobody wrote it for this system. Every case received a reading. On 645judged cases it governed 92.6% of the time, and inside that governing produced138 distinct configurations of concern. Across an exhaustive enumeration of57,344 structural actions it governed 95.9% of the time across 12,236 distinctlevels of concern. A control that allows or blocks resolves two. Where aconventional severity ordering says one action outranks another, the systeminverts it, because it weights standing and consent against the character ofthe act rather than summing stakes. The paper reports what the run did not establish as fully as what it did: apre-registered stop-gate fired and was honored, a pre-registered hypothesis wascontradicted, and the model-mediated reading of prose into structure agreesacross vendors only 6.6% of the time. This record contains the paper and its reproducibility package: aggregatetables covering the posture distribution, the 138 concern configurations underopaque identifiers, the concern-mass distributions, the cross-vendorcontingency table, the re-gate counts, and the unread-field distribution,together with a script that derives every reported figure from those tablesalone. The underlying per-case records and the specification of the moral basisare not deposited and are available under agreement.