Frame-type diversity
The variety of frame types in play — the kinds of framing a region of the substrate is holding, not just how much activation it carries — is the most direct read-out of where a population sits in the divergence/convergence cycle. Exploration widens the set of types; convergence narrows it deliberately; the failures are a set that stays narrow when the task still needs range, or never settles.
Under sustained write-back the failure has a characteristic shape. Retrieval concentrates on what already dominates, so the variety of types narrows on its own: homogenization drift, the substrate-level analogue of model collapse, which the literature has begun to measure in collective AI output (Doshi & Hauser, 2024; Wenger & Kenett, 2026). It is the same shape as a belief attractor — the more context that settles around a region, the harder it is for new material to compete.
How varied a set actually is — and how to read it across scales — is a measurement question, handled by Hill numbers. Whether a given level of variety is appropriate depends on the task, which is the open part: see task-appropriate behavior.