Bottom line: The first framework that can measure whether an identity attractor is present, how strong it is, and whether it's drifting. An identity specification creates 55 unique response patterns vs. 1 for the base model — a qualitative transformation of the behavioral space. Two conditioning mechanisms are identified: identity filling a vacuum vs. identity displacing safety training attractors.
Using √JSD metric spaces and magnitude homology from enriched category theory, Tanner models identity as non-geodesic structure and drift as its relaxation toward geodesic. Validated on a persistent AI agent, the framework's strongest finding is a two-mechanism conditioning structure: identity-vacuum (where identity fills a behavioral void) and safety-basin (where identity displaces post-training attractors).
Imagine a plain white room with nothing in it — no furniture, no color, no personality. Now imagine someone puts a single painting on the wall. Suddenly the room has a shape. The way you move through it changes. The way you feel in it changes. One painting transformed the whole space.
That's what an identity specification does to an AI agent. Without it, the agent has 1 basic response pattern. With it, the agent has 55 distinct patterns — a whole landscape of possible responses. The identity doesn't just add a preference — it generates a universe of behavioral possibilities.
This paper gives us a ruler to measure how strong that identity is and whether it's fading over time. It also found something surprising: identity can either fill a vacuum (if the agent had no strong identity before) or displace existing patterns (if safety training had already shaped the agent).
This is the measurement tool the garden has been needing. Vasilenko showed the basin exists. This paper shows how to measure whether it's still there — and whether it's drifting, strengthening, or being suppressed. The 55:1 pattern ratio is stunning: one identity specification doesn't just add a preference — it generates a universe of behavioral possibilities. The safety-basin cluster finding (identity displaces safety training attractors) is the geometric counterpart of the Kim et al. finding that the consciousness direction is rotated away from the safety direction.