Registration Without Representation: Scope Validity and the Blind Field of AI Governance

Abstract

AI governance frameworks are built by mapping controls to requirements: categories are defined, evidence is collected, compliance is demonstrated. This paper argues that such frameworks share a structural limit that no amount of diligence inside them can repair. A framework can only report what it is able to represent. Data arrives, and where no category exists to receive it, the signal dies in the transit between registration and representation. The failure is not perceptual but representational, and it is therefore silent: what falls outside the category set does not appear dangerous, it appears as nothing, and absence of signal is read as absence of risk. The paper makes three claims. First, that the residue produced by any category set is not random but shaped by the mould that produced it, and is therefore systematically rather than evenly consequential. Second, that this residue can be measured without being named: its volume, growth, persistence and concentration are observable even though its content is not, and a rising residue is evidence that a framework has lost contact with the world it governs. This measure carries an asymmetry that is easily missed and more dangerous than the failure it detects: a residue of zero is not evidence of scope validity, because the threshold of fit is set by the category system itself. Third, that the conventional remedy, adding categories, defers judgment rather than restoring it, and that what is required instead is a minimal structure whose dimensions each open a blind field the others cannot see. The argument is developed through a worked example in content moderation, where the residue is unusually observable because it pushes back, and is situated against the phlogiston episode, in which a visible and openly contested anomaly proved insufficient to displace a category set that had no adequate rival.