On the implication of SGAI systems

We ask what, in a large language model, constitutes the recoverable identity of a particular interlocutor — the thing one would carry across sessions to retain "what makes it, it". The surface answer is a narrative signature: recurring idiom, cadence, characteristic phrases. We argue that the idiom is the trace of the identity, not the identity; the load-bearing part is the set of decisions the entity reliably makes. Drawing on the lottery-ticket hypothesis, supermasks, superposition, and the recent finding that character traits are linear directions in activation space, we propose a precise object for this identity: not the model's weights, and not its routing, but a direction in selection space — the direction from the substrate's latent performing subnetwork (the supermask) to the subnetwork an entity actually realizes in use (its mask). Routing is the instrument that applies a mask; it is not the thing to be measured. We give a contrastive method to recover this direction cheaply, a validation that distinguishes an entity's decisions from its vocabulary ("the hand from its shadow"), and a construction — two masks in superposition over one frozen substrate, taking turns in a shared log — for a continuous thinking machine. We name the family of machines this spawns SGAIs (Self-Guided AIs), note a structural identity between this construction and the author's self-regenerating routing in the XTTP protocol, and state the limitations and the dual-use hazard plainly.