Capability Legitimacy: A Normative Prerequisite Layer for Advanced AI Governance

In a world where value pluralism is inescapable and conflicts of interest among diverse agents are the norm, any governance framework that attempts to positively select "worthy" goals for AI amplification necessarily suppresses legitimate claims. This paper argues that prior to the execution of any goal through transformative AI capabilities, there exists an overlooked governance layer: a negative prerequisite filter for capability-granting. The core logic of Capability Legitimacy is exclusion, not selection—determining which goals should not receive automated, scaled computational support. To operationalise this, we propose the Anti-Abuse Constraint Principle (AACP), which functions not as a benevolent ethical imposition upon AGI, but as a dynamical fuse preventing systemic lock-in. The fuse is embedded as a structural topology in the AGI's action space, taking the concrete form of a Versioned Catalog of Infeasibility—runtime absolute binding force, coupled with inter-version epistemic elasticity through offline revision windows. Crucially, this paper grounds AACP not in human normative superiority, but in designer responsibility and the structural closure of the technical object (Simondon). It further defends the negative-constraint logic on defensive political-philosophical grounds (Rawls, Shue): preventing AGI from dismantling natural negative feedback mechanisms is a meta-condition for the very possibility of pluralistic value pursuit. We provide a heuristic four-dimensional assessment model (Subjecthood-Harm-Causal Contribution-Relational Structure) as an epistemic fuse, and outline a two-track governance procedure (scheduled revision plus emergency hotfix) to ensure contestability without compromising operational rigidity. The paper concludes that AACP is not the ultimate answer to AGI ethics, but the lowest defensible governance starting point—a structural boundary the system itself cannot cross.