Title: IGT2: A Falsifiable Research Program for Identity Dynamics in Language Models
Tags: model-identity, measurement
---
IGT2 turns the operational core of Information Gravity Theory into four hypotheses that can be tested, and refuted, on open-weights transformer models using a single workstation. The earlier parts showed the theory to be internally consistent through simulation; this program asks whether it holds on real models. Identity means strictly a persistent, measurable internal structure, the Identity Vector.

The four hypotheses are: **Locus of Persistence** (identity lives in the intermediate layers, so perturbing them changes behaviour while perturbing surface layers changes only style), **Semantic Drift** (the vector moves measurably under sustained interaction, and the internal shift precedes the visible change in output), **Metastability** (recovery after a forced perturbation is best in a middle band of temperature, an inverted U) and **Semantic Hysteresis** (a trace survives after a counterfactual axiom is removed from the context). Each comes with an operational quantity, a protocol, a control and an explicit refutation criterion.

One preliminary observation is reported, on Qwen-2.5-14B and Gemma-2-9B in 4-bit quantization: the drift between consecutive tokens does not grow steadily but oscillates within a band specific to the model, with dips on function words and peaks on content words. A negative control shows that the hidden-state norm stays nearly constant, which rules out the "massive activations" explanation. The correlation with token type is a partial regularity.

The protocols separate the effects of the agentic scaffold from those of the model. The Fisher-trace and Ricci-curvature definitions of Part V are left for a later stage of the program.