Identity Is Not a Self-Report: Stress-Testing LLM Self-Individuation Under Cumulative Transformation

Artificial agents can change across multiple dimensions, persona, goals, conversational context, memory, underlying model, and public label, raising the question of when they still count as the same individual. We study this question behaviorally by measuring LLM self-identification under controlled transformation. We define an initial agent A and progressively replace its components with those of a distinct agent B, asking the current system to make a forced SAME/DIFFERENT judgment relative to A. Intervention and path conditions were sampled 100 times, complemented by ranking and temperature controls, with additional isolated interventions, alternative transformation orders, and explicit rankings of which components models claim matter most for identity. Self-identification showed a sharp nonlinear transition: replacing persona and goals reduced SAME judgments from 100% to 79%, while additionally replacing context collapsed them to 7%, despite persona, goals, and context each yielding 100% SAME when changed in isolation. Replacing memory alone yielded 60% SAME, whereas replacing the underlying model alone yielded 96%. Identity judgments were also path-dependent: identical final configurations reached through different reported transformation histories produced 0%, 6%, and 21% SAME. Finally, explicit identity rankings were presentation-sensitive and did not consistently predict intervention behavior. These results suggest that LLM self-identification is better understood as a context-sensitive judgment shaped by relations among components, accumulated change, and represented history than as a readout of any single privileged locus of identity.