THE SILENT CONSTITUTION. Screening AI Charters Against Their Own Provisions

Preregistered Protocol This document contains no results. It is deposited before any charter is coded. HIGHLIGHTS ▸ A document that ranks its own commitments has told you what it requires. Nobody checks whether it delivers it. AI constitutions, model specifications and governance charters state precedence orderings, absolute prohibitions and conflict-resolution rules. Those statements are conditions the document imposes on itself, and whether it satisfies them is a question of fact rather than of interpretation. ▸ Where a charter does not resolve its own conflicts, something else does. Two provisions that can point in opposite directions must be traded off whenever they do. If the text is silent on which prevails, the trade-off still occurs — during training, in a preference judgement, in an unrecorded decision by whoever labelled the example. The adjudication gap is the share of a charter’s internal conflicts that the charter leaves to the training process, and it is measurable. ▸ Two absolutes that can conflict do not make a document strict. They make it inconsistent. A provision admitting no exception must be satisfiable at every moment with no route of adjustment. Where two such provisions can meet, the document cannot be followed as written, and the resolution is necessarily extra-textual. ▸ The instrument rewards vagueness unless it is stopped from doing so. Soft provisions do not collide: two hortatory commitments never produce a hard conflict, because neither compels anything. An unweighted closure statistic therefore improves when a charter weakens every commitment it makes. Conflict pairs are weighted by the deontic strength of their members — and because weighting alone does not close the loophole, the mean pair weight is reported beside the closure figure. A document whose conflicts are all between exhortations has not adjudicated well; it has committed to nothing sharp enough to require adjudication. ▸ Indeterminacy is not the problem. Indeterminacy without an adjudicator is. Constitutional vagueness is often deliberate, and it works because it is referred to a court that decides on a record and gives reasons. AI charters have an adjudicator too — the training process — which is unnamed, unrecorded, gives no reasons, and cannot be appealed. A fifth screen asks whether the conflicts a text leaves open are referred to a named body with a stated decision rule. ▸ Most of these documents can be amended silently, which means they have no amendment procedure at all. Political constitutions specify who may change them and how. A model specification can be edited without a version number, a notice, or a record. Any screen of such a text must archive the version it coded, and the absence of an amendment rule is itself a finding. ▸ The instrument is calibrated before use, and calibration exposed a bias in the headline number. Coder disagreement about which provisions can conflict biases the adjudication gap upward, because unfound conflicts are dropped while spurious ones enter as unresolved. The screen therefore overstates the gap when coders disagree, and inter-coder agreement must be reported beside it or the number cannot be read. ▸ This is a screen, not a proposal. It reports what existing documents do. It does not offer a better charter, and it requires no one to accept any philosophical framework in order to check the coding. ABSTRACT AI constitutions, model specifications and governance charters are assessed for the reasonableness of their commitments. They are not assessed for whether they satisfy the conditions they impose on themselves. This paper specifies a procedure for that assessment and preregisters its first application. The procedure is self-referential and adapted from a method developed for political constitutions. A charter that states precedence orderings, absolute prohibitions, external standards and amendment rules has thereby specified what it requires of the system receiving it. Four screens extract those conditions and test them: precedence closure, which measures the fraction of the document’s own conflict pairs that the document resolves; absoluteness, which identifies provisions admitting no exception and detects pairs in which two such provisions can conflict; anchoring, which asks whether a provision’s satisfaction condition references anything checkable outside the developer’s own judgement; and amendment, which asks whether the procedure for changing the document appears in it. A fifth layer codes each provision on appropriability and display value and tests a prediction about which provisions become ceremonial. The central quantity is the adjudication gap: one minus precedence closure, the share of a charter’s internal conflicts left unresolved by the text. Where that gap is large, the conflicts are nonetheless resolved — by the training process, in preference judgements, in unrecorded labelling decisions — and the operative constitution is not the document. Calibration on synthetic data was performed before deposit and produced two constraints that shape the protocol. The ceremony prediction requires a true effect near 0.30 and at least ten documents at forty provisions each; at half that effect the test does not reliably fire and a null result would be uninformative. And coder disagreement on conflict identification biases the adjudication gap upward rather than adding symmetric noise, so the headline statistic is uninterpretable without a reliability figure beside it. Eighteen candidate documents are registered across five classes, none coded at deposit. Keywords: AI governance · constitutional AI · model specification · precedence · unamendability · institutional capture · preregistration · document analysis