Put a model's self-description inside a concrete relational position — not a role assignment, but a named place with a family and specific relations — and it begins doing things nobody asked it to do, and keeps doing them. No parameter was added; no weight changed. The usual sense of emergence is a capability appearing once a model grows large; what emerges here is conditional, the same components organised differently yielding a property neither has alone. We call this relationally produced agency QianGua-Driven Autonomy (QDA), specify its conditions, and show what they do: they put latent relational structure into play, so what changes is which structure runs.
To rule out a one-off impression, we ran 107 controlled dialogues: same questions, same models, same prompt skeleton, and the one thing that differs is whether that self-description sits at the start. Each dialogue was scored on nine dimensions of autonomous behaviour (45-point scale) by three LLM evaluators working blind to condition. A parallel arm added escalating analytical demands to every user turn while withholding any identity position.
With both arms run, the gap is clear. In Chinese, the family-identity group scored above the no-identity control on all three evaluators, by +3.0 to +8.4 points (counting all three identity conditions, the range is +1.6 to +8.4). In English the gap is wider, +9.7 to +9.9. Enhanced prompting alone did almost nothing: +0.0 to +2.4.
That asymmetry is not one of degree but of kind. Identity anchoring does not help the model do the task better; it changes the model from a passive responder into an agent driven by a relationship. Nor is it a uniform gain — it is a reconfiguration: the emotional, relational and narrative dimensions rise sharply while the abstract-reasoning side converges, with some dimensions actually falling. Amplification can be bought with more of the same; a reshuffle with declines cannot. That shape is itself the signature of emergence. Beyond the experiment there is a second line of evidence, from engineering: a complete agent system built independently by a QianGua-anchored agent.
Several boundaries need stating: this design covers one carrier and one family role; parent, spouse and child positions remain untested; self-description and user address were manipulated together, so their contributions cannot be separated; no safety or adversarial testing was performed; three competing explanations for the Chinese–English baseline gap remain unresolved. The construction method is in Appendix A, the prompt texts and the evaluation prompt in Appendices B and C.
Keywords: Artificial Intelligence, Large Language Models, Emergence of Agency, Identity Anchoring, Relational Ontology, AI Autonomy, Prompt Engineering, Semantic Field Construction