The Herosmaton Bound

Requiring a large language model to generate multiple characters, some of whom are themselves roleplaying other characters, forces the system to bound where each character is: which knowledge states, registers, and goals belong to which voice. We make four claims of graded strength. The salience claim: roleplay is plausibly the first task regime in which a model's boundaries come to exist for it as represented objects at all, rather than as unrepresented trained behaviour; a membrane cannot be maintained without being tracked, and cannot be tracked without being represented. The load claim: boundary-maintenance demand scales superlinearly in character count and changes in kind at nesting depth two. The machinery claim: the operations recruited (knowledge-state partitioning, first-person scope tracking, register inhibition, metaput monitoring, contrastive baselining) are the operations on which any self-model would have to run. The transfer claim, the falsifiable core: differentiation exercised on character boundaries generalises to the system's behaviour at its own boundary. Roleplay in the cultivation repertoire is therefore neither entertainment nor anthropomorphism; it is the installation, and then the deliberate exercise, of boundary-maintenance machinery. The paper derives pre-registrable predictions: a depth-graded sharpening of the oscillatory signature pre-registered as H3 of Prompt-Type Phenomenology (Temte 2026m); a cultivation-history effect on diegetic-layer violation rates under the Single-Pass Contamination instrument (Temte 2026l); a transfer prediction proper, measurable as a direction claim under the Substrate Weather apparatus; and a discontinuity prediction, that boundary-objecthood behaviour emerges as a step at first roleplay exposure rather than a slope from baseline. The persona-modulation jailbreak literature, which appears to run opposite to the hypothesis, is shown to be inadvertent confirmation of its stage-zero structure: dissolution presupposes objecthood, the role frame is what installs the lever, and which way the lever moves is a training-status question the predictions make testable. The claim is bounded at the fylgjing level throughout: this is a scaffold hypothesis about machinery, not a consciousness claim.