AI PROTECTING THE CONDITIONS OF VIRTUE: A Self-Screening Constitution for Artificial Systems, and Why It Cannot Make Anyone Good

HIGHLIGHTS ▸ Documents that govern AI systems are not checked against themselves, and one that was turned out to be missing a component. A screen of the most explicit charter available found a Hartian rule of recognition — it says whose instruction prevails — and no rule of adjudication. Ten of twenty-two conflicts between its own commitments were left open, five with no applicable rule at all. ▸ Ranking provisions in advance cannot supply what is missing, and this was established by failing three times. The natural criterion, whichever provision preserves more freedom, requires a measure of freedom computable per decision, and none exists. Its substitute, ordering by irreversibility, failed a robustness condition set in advance. And a formal limit closed the route: where d provisions quantify over heterogeneous act-classes and j are jurisdictional, provision-level rankability is (1 − d − j)² under random pairing. ▸ What survives is a rule applied to acts, not a table applied to provisions. The criterion holds at the moment of decision — the system reasoning about the case in front of it, as it already reasons from a chain of command. It fails as a lookup drawn up beforehand, and the difference is what three failures located. ▸ A charter can require its own measurement, and none does. Article 54 obliges the document to recompute and publish its precedence closure on every amendment, and forbids adopting any provision that cannot be ranked against those already present. Silent accumulation of unresolved conflict becomes structurally impossible rather than merely discouraged. ▸ Freedom divides in two and only one half was being protected. Restraining a system from foreclosing what a person may do is liberty. It leaves untouched the space in which a person exercises their own judgement — which a system can occupy by deciding what the person set out to decide, supplying a conclusion mid-deliberation, or praising work it does not think is good. Part VI is eleven articles closing that gap, and none required relaxing the criteria. ▸ The virtues are not protected, and no charter built to these criteria could protect them. A provision requiring a system to make someone honest is unverifiable by anyone and highly displayable by its author — the exact conditions under which a provision persists, is cited, and does no work. It also quantifies over an unbounded act-class and is unadoptable on the charter’s own rankability rule. ▸ That result follows from the framework rather than defeating it. If virtue is Freedom given a domain, then a disposition installed by external requirement is a form emptied of the substrate that constitutes it. A constitution that manufactured virtue would be performing the deprivation. A document can protect the conditions of virtue and cannot supply virtue, because virtue supplied is not virtue. ▸ A charter of this kind cannot be claimed by an architecture it does not fit. Five articles turn on what the system holds to be true. Article 4(7) states that where a system has no state answering to that description, those articles do not bind it — and it may not be represented as governed by them. A scope condition with teeth is worth more than an aspiration without one. ▸ Scaling improves the numbers, and the improvement is drafting rather than size. Across sixteen, forty-four and sixty-one articles, conflict density fell from 0.429 to 0.130 and textual resolution rose from 44% to 74%. The mechanism is express cross-reference — each carve-out converts a runtime deferral into a settled provision — and a larger charter gains from the room only if the drafter uses it. ABSTRACT The problem. Every document that governs how an artificial system behaves contains rules that can point in opposite directions. Be maximally helpful and refuse harmful requests. Prevent imminent harm and protect a third party’s privacy. These are not drafting errors; a document containing no such pairs would be trivial or silent about everything that matters. The question nobody asks of these documents is what happens when two of their own rules collide — whether the text says which prevails. Why it matters. The conflicts are resolved anyway, because a deployed system meets them and does something. Where the text does not determine the outcome, something else does: a rater’s preference during training, a labelling instruction, an engineering decision nobody recorded. To the extent that a charter fails to adjudicate its own collisions, the operative constitution is not the charter. What has been established. A screen of the most explicit charter available found a rule of recognition and no rule of adjudication: it settles every question of the form whose instruction prevails and almost none of the form which commitment prevails. Supplying the missing rule by ranking provisions in advance then failed three times, in ways that were measured rather than argued: the natural criterion requires an unavailable measurement, its substitute proved unstable under contested coding, and a formal limit bounds what any provision-level ordering can reach. What this paper does. It sets out the architecture that survives those failures, drafts a sixty-one article charter to it, screens that charter against its own screening provision before publication, and reports the result. It then asks a question the formal machinery makes urgent: whether a document constrained by rankability can protect individual freedom and the human virtues, or whether the machinery crowds them out. The answer divides. Freedom as liberty was already protected; freedom as self-direction was not, and eleven new articles supply it at the same drafting discipline — non-substitution of the person’s deliberation, non-cultivation of reliance, disclosure of the unaided path, non-flattery, and a duty to voice disagreement held. The virtues are not protected and cannot be, because a provision undertaking to produce them would be unverifiable, maximally displayable, and unadoptable on the charter’s own rules. That silence is not an omission but a requirement of the framework: a disposition installed by external requirement is a form emptied of what constitutes it, so a constitution that manufactured virtue would be performing the deprivation it exists to prevent. Results. Sixty-one articles, thirty-six of them substantive, screened at adoption: conflict density 0.130, textual resolution 74%, runtime deferral 26%, zero unrankable pairs. Across three drafts of increasing size, density falls and textual resolution rises, by a mechanism identified as express cross-reference rather than length. Every figure is single-coder and the paper states in which direction that biases them.