WHEN AI CONSTITUTIONS SPEAK AGAINST THEMSELVES: The First Charter Screen, Measuring Whether AI Governance Documents Resolve Their Own Conflicts

ABSTRACT The problem. Every document that governs how an AI system behaves contains rules that can point in opposite directions. Be maximally helpful and refuse harmful requests. Prevent imminent harm and protect people’s privacy. These are not drafting errors; a document containing no such pairs would be either 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 one wins. Why it matters. The conflicts do not go unresolved in practice. A deployed assistant meets them constantly and does something. If the document 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. This is measurable, and until now it has not been measured. What this paper does. It reports the first application of a screening procedure adapted from comparative constitutional law: extract from a document the conditions the document imposes on itself, then check whether it satisfies them. Two texts were coded — the root-level sections of the OpenAI Model Spec dated 18 August 2026, and the core normative articles of the European Commission’s 2021 proposal for an Artificial Intelligence Act. What it found. The Model Spec adjudicates far more than a general scepticism about corporate charters would predict, and it adjudicates the wrong thing. Its chain of command is an authority ordering over speakers — root outranks system, system outranks developer, developer outranks user — and it settles every question of the form whose instruction prevails. It does not rank the document’s own commitments against each other, and it cannot, because both speak with the same voice. Ten of twenty-two identified root-level conflicts remain open, five of them with no applicable rule at all. In Hartian terms the Spec has a rule of recognition and no rule of adjudication, and its stated commitment to continuous updating without a procedure means it has no rule of change either: one secondary rule out of three. Two measures, and the distance between them. A conflict can be resolved in two different senses. π_RANK asks whether the text states which provision prevails; π_ACT asks whether it states what to do. The Model Spec returns 0.586 and 0.667, and the 0.081 between them is silent adjudication — behaviour fixed by a rule that never states the ranking that fixed it. The AI Act proposal returns 0.833 for both, a gap of exactly zero, which is what a drafting tradition that ranks provisions explicitly looks like when measured. No adjudicator where it is needed. Full adjudicator coverage — a named body, a stated rule, available at the moment the conflict arises — is zero in both documents. Post-hoc mechanisms exist and are credited, reaching 0.400 and 1.000. Neither text directs the assistant to invoke one when a collision occurs, which is the claim the screen can support: not that real-time escalation is impossible, but that these documents do not provide for it. On the reliability of these numbers. Two documents were coded by one person, who is the author of the instrument, and no inter-coder estimate is available or claimed. In place of one, the paper reports the interval within which the headline figure must lie under any recoding of the judgements this coder found contestable. For the Model Spec that interval is [0.586, 0.798]. The adjudication gap of 0.414 is therefore an upper bound: a second coder can shrink it and cannot enlarge it. The pilot also produced four corrections to the instrument and one to its operator, all recorded. Keywords: AI governance · model specification · chain of command · precedence · adjudication gap · silent adjudication · rule of