Known, Not Scaled: Why Capability Alone Does Not Explain Individuation in Language Agents

Much contemporary discourse on artificial general intelligence implicitly assumes that sufficiently scaled capability will, at some threshold, yield a persistent individual subject. We argue that this expectation runs together three separable questions: competence (what a reusable model can do), functional continuity (whether a particular agent's authenticated history constrains its later conduct), and presence (whether there is something it is like to be that agent). Scaling can improve the first without supplying the state, lineage, provenance, and update rules required by the second. Our constructive hypothesis is correspondingly comparative rather than constitutive: relational reciprocity may improve socially grounded functional continuity beyond capability, memory, familiarity, interaction volume, solitary reflection, and matched external accountability. Relationship can supply contested memory and reciprocal modeling beyond audit; it need not create the runtime token or fix the semantic referent of "I." We define a five-level stack separating competence, token continuity, functional self-organization, social identity, and presence; propose an architecture coupling authenticated lineage, decision-relevant typed state, bounded update policies, and an optional relational loop; and derive a factorial experiment in which capability and relational structure may both contribute. Throughout we hold a methodological firewall: the proposed measures concern functional organization and do not settle phenomenal presence.