Peter Nausner

Agent-safety infrastructure is beginning to move the enforcement of limits out of the AI agent, onto separate hardware and into monitors the agent cannot get past. This research note argues that the word authority, as the debate uses it, covers two things that are now moving separately: enforcement, which makes a limit hold, and setting, which determines where the limit lies. A profile of outside instances measures them apart, among other dimensions by externality to the agent and to the operator, by power to enforce, by power to set, and by binding effect. Applied to six cases from September 2026, from embedded evaluation and cross-developer incident reports to California's verification statutes and NVIDIA's open agent-safety reference design, it shows that observation is externalised widely, that enforcement leaves the agent but not the operator, and that outside setting exists for verifiers rather than for agents. Proposals to limit the production of capability reproduce the pattern. A bill adopted by the Austrian Parliament, proposed § 18a of the General Administrative Procedure Act on fully automated administrative acts (not yet promulgated as of 12 October 2026), shows what operator-independent grounding, visible deployment and review can look like, and where they stop. The note derives three requirements for limits that affect third parties, and a boundary: a system may enforce a limit that someone has set; it must not set a limit that no one has. A postscript covers developments to 10 October 2026, including a New York City Council bill on third-party validation and shut-down capability. Note on tools: The author used AI-assisted drafting, translation, and editorial tools during the preparation of this manuscript. The conceptual framing, theoretical argument, selection of sources, responsibility for the content, and final editorial decisions remain with the author.

Most evaluation of artificial intelligence measures task competence, that is, performance on benchmarks posed within a single space of problems. We argue that general intelligence in delegated, institutionally embedded contexts is instead transformation competence. A capable system must occupy a plurality of reason-spaces, each a vocabulary with its own material implications and incompatibilities. It must transform reasons correctly as they cross the boundaries between these spaces, preserving an identifiable entity's reason-relational profile where the boundary permits and refusing transfer where it would violate a defining incompatibility. And it must grow a new reason-space where transformation across the existing ones systematically fails. We call the core norm anaphoric fidelity. A transformation that claims to track the same entity must preserve enough of that entity's reason-relational profile for the claim to hold, and breaking it is a covert referential substitution. Monolithic scaling cannot yield transformation competence, because by collapsing the plurality into one space it can only transfer a reason flat, treating it as valid across boundaries without transformation. Flat transfer is the fault behind familiar failures, from an assessment that drops a person's standing at an institutional boundary to an agent that carries an instruction or a permission across a context boundary, the confused deputy. We ground reason-spaces in implication-space semantics, give a typology of what can happen to a reason at a boundary, illustrate the criterion on institutional transitions in disability policy, and propose a transformation-competence benchmark that operationalizes both the criterion and the anti-monolith claim. The criterion has two faces. Over entities it is a criterion of competence, testable on a system. Over persons it is a criterion of justice, identified with legal capacity. Note on tools: The author used AI-assisted drafting, translation, and editorial tools during the preparation of this manuscript. The conceptual framing, theoretical argument, selection of sources, responsibility for the content, and final editorial decisions remain with the author.

A polycontextural agent architecture keeps several validity logics apart rather than collapsing them into one. Such an architecture faces a learning problem of its own. Updating model weights is the wrong form of improvement here. It is opaque, and it tends to homogenize a plurality of perspectives toward a single voice. We ask whether such an architecture can instead learn by changing its validity architecture, and in particular whether it can grow a new validity space when its existing ones prove inadequate. We make four contributions. First, we characterize a content-blind reflection operator by which a contexture, shown only that it is in conflict and not the content of the conflict, develops new aspects of its own logic without importing the other’s. Second, we describe how curated records of these aspects re-enter later runs as non-binding prior findings, so the architecture learns across cases without weight updates. Both mechanisms are implemented. Third, and centrally, we propose a criterion for the genesis of a new contexture. A new validity space is born when a recurring conflict can be addressed only by violating an import prohibition, that is, when a recurring content has no place that can legitimately house it. We ground this as an inferentialist genesis of logical places under a polycontextural constraint, using Brandom’s mechanism of conceptual genesis against Brandom’s own holism, with Günther’s plurality as the constraint. Fourth, we describe how a newly born contexture matures, and when it may take part in the birth of another. The genesis and maturation mechanisms are specified rather than implemented, and we present them as a criterion, its justification, and a build program, illustrated on a social-policy demonstrator. The result reframes learning in delegated AI as change to an auditable validity architecture rather than to a model’s weights.

Current debates on artificial general intelligence often imagine AGI as the emergence of a unified artificial mind. This paper questions that monolithic assumption. It does not ask whether AI systems possess phenomenal consciousness. It asks whether increasingly capable systems can distinguish, occupy, and mediate heterogeneous validity logics in real decision contexts. Starting from the definitional heterogeneity of the AGI debate, the paper argues that general intelligence is constituted across multiple forms of validity, relevance, and responsibility. Building on Gotthard Günther's theory of polycontexturality, it conceptualizes these heterogeneous logics as contextures: distinct validity spaces in which claims, recommendations, and actions become meaningful under different conditions. The paper develops the concepts of contextural reflexivity, delegated AI, and contextural accountability. It proposes polycontextural agent architectures as an alternative to monolithic AGI models. Such architectures combine contexture-specific agents, mediation layers, role structures, and governance mechanisms in order to make explicit which validity logic an AI-generated claim follows, which other contextures it affects, and when human or institutional escalation is required. The argument is illustrated through a demonstrator from inclusive labor-market integration. The contribution is a conceptual framework for AGI theory, agentic AI, alignment, and AI governance. It reframes general artificial intelligence as a problem of reflexive mediation within delegated and institutionally embedded decision architectures. The conclusion outlines three subsequent research directions: Organizational Validity Mapping, Polycontextural Compliance Metrics, and architectural learning through transjunctional reflection.