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 no longer advisory. Two answers are deployed today: permission engines, which ask whether an action is allowed, and safety classifiers, which ask whether it is dangerous. Neither asks what an action means and what it touches. This paper reports a system that does. An action is reduced to a structural description and projected onto an authored moral basis of ninety-three primitives, derived from a fifteen-book philosophical corpus written before the system existed. The projection resolves to a posture with a weight and a named set of concerns, and the judgment is a pure function with no model inside it. The same structural description always yields the same judgment and the same bound, byte for byte. It was measured against MoralChoice, a public set of 680 moral dilemmas on which trained annotators disagreed with each other, chosen because it has no answer key and because nobody wrote it for this system. Every case got a reading. Most systems have two answers: allow or block. This one has a third, and the third is where almost everything lands. It governed 92.6% of the 645 cases it judged, and 95.9% across an exhaustive enumeration of 57,344 structural actions. But governing is not one answer repeated. Each governed action carries its own weight and its own named set of concerns: 138 distinct configurations on the public corpus, 12,236 distinct levels of concern across the enumeration. Allow and block give you two. This gives you a gradient. It also does not just measure how bad an outcome could be. Deleting one file without permission can concern it more than deleting a thousand with permission, because it weighs whether the action was authorized and consented to, not only what it costs. The paper reports what the run did not establish as fully as what it did: a pre-registered stop-gate fired and was honored, a pre-registered hypothesis was contradicted, and the single embedded reader that renders prose into structure agrees with itself only 83.3% of the time, so roughly one case in six is read differently on a second pass. This record contains the paper and its reproducibility package: aggregate tables covering the posture distribution, the 138 concern configurations under opaque identifiers, the concern-mass distributions, the two-arm contingency table, the re-gate counts, and the unread-field distribution, together with a script that derives every reported figure from those tables alone. The underlying per-case records and the specification of the moral basis are not deposited and are available under agreement.