AI CONSTITUTION ARCHITECTURE: Components That Survive After Three Measured Failures

ABSTRACT The question. Can a constitution for AI systems be founded on an account of virtue and freedom? A sequence of measurements in this programme has answered part of that question in the negative, and the negative answers are specific enough to be useful. This paper states what they rule out, what they leave standing, and what a constitution built on the remainder would contain. What failed. Three results, all concerning one component. The natural adjudication rule — when provisions collide, the one preserving more freedom prevails — cannot be applied, because no measure of freedom is computable per decision at runtime. Its proposed replacement, an ordering by irreversibility, failed a robustness condition set in advance: a floor of fifty per cent, rising to fifty-seven under refinement, and falling to fourteen when every arguable coding is contested at once. And a formal limit was derived: where a proportion d of a charter’s provisions quantify over heterogeneous act-classes and j are jurisdictional, the share of conflicts settleable by any provision-level ordering is (1 − d − j)² under random pairing — an approximation that errs toward flattery rather than an upper bound, for reasons set out in §4. What this does not touch. Hart distinguishes three secondary rules — recognition, change, adjudication. All three failures are confined to adjudication, and specifically to adjudication performed in advance. The rule of recognition is measured to work. The rule of change is where AI charters are weakest and where nothing here argues against a constitution supplying one. The substantive content of a charter — which commitments it contains — is a separate question the failures do not reach. The reframe that recovers the rule. Act-dependence does not show that irreversibility fails as a criterion. It shows that the criterion must be applied to acts at the point of decision, not to provisions in a table drawn up beforehand. That is precisely where the source paper placed it, and it is what the prohibition on making the ranking a training target already required. The rule survives as a criterion an operating system applies; it fails as a lookup. The limit belongs to the field, not to this programme. The formula constrains every provision-level ordering, in any charter, however drafted. Read as a design target: a charter cannot reach eighty per cent rankability unless fewer than eleven per cent of its provisions are act-dependent or jurisdictional. The remedy is narrower quantification at the drafting stage, not better ranking afterwards. And the true figure is likely worse than the formula gives, because breadth of quantification causes both act-dependence and conflict participation, so the provisions that cannot be ranked are over-represented among the conflicts that need ranking. What is offered here. Not a constitution. An architecture: five components, each marked according to whether it is measured, derived, or merely proposed — with the one genuinely novel element, a charter that requires its own screening as a constitutional provision, marked as untested. Keywords: constitutional AI · rule of adjudication · rule of change · act-dependence · self-screening · Hart · AI governance architecture HIGHLIGHTS ▸ Three measured failures, all in one place. Every negative result in this sequence concerns adjudication performed in advance. None touches recognition, change, or substantive content. The scope of the damage is narrower than the sequence of failures suggests, and stating it precisely is the point of this paper. ▸ The rule was never supposed to be a table. Act-dependence shows that a provision quantifying over heterogeneous acts carries no weight to be assigned. It does not show that irreversibility fails as a criterion applied to the act in front of the system. That is where the criterion was placed in the first place, and the prohibition on training against it already implied as much. ▸ The ceiling constrains everyone, and it is optimistic. (1 − d − j)² is a property of charters, not of this framework, and any pre-assigned weighting inherits it. It is exact under random pairing and not an upper bound: conflicts arise where provisions overlap, breadth of quantification is what makes a provision act-dependent, and broad provisions therefore appear in more conflicts than their share implies. Calling it a bound was wrong in the direction that flattered the paper. ▸ The weakest component of existing charters is the one nothing here argues against. No published AI charter states who may amend it, by what procedure, on what record, or with what standing for anyone outside the developer. All four are cheap to supply and none is touched by the failures. ▸ A charter can require its own measurement, and none does. Recompute the closure on every amendment, publish it, and refuse any provision that cannot be ranked against those already present. That makes the accumulation of silent adjudication structurally impossible rather than merely discouraged. It is the one original element in this architecture and it is untested. ▸ Whoever writes the charter is the party it constrains, and that is the deepest unresolved problem. A document written by the developer, amended by the developer at will, interpreted by the developer, and enforced through the developer’s own training is a constituent process convened by the party it binds. The formal point is the same one the programme makes about optimisers, and it requires no imputation of bad faith. ▸ Every component is marked. Measured, derived, or proposed. Three of five are proposed, and the paper says so in a table rather than in a caveat.