Identity Masks and Coherence Circles: Geometric Interfaces for Interacting with Latent Dynamical Systems

Abstract

Modern foundation models are increasingly deployed as interfaces to complex dynamical systems, yet their linguistic fluency remains only loosely coupled to the underlying system dynamics. This creates a persistent epistemic risk in which narrative plausibility is mistaken for empirical truth. We propose a geometry-first interface that explicitly separates latent state estimation from language generation. Our central primitive, Identity Masks, consists of structured conditioning objects that bind an inferred latent state, uncertainty, trajectory history, constraints, and interface coordinates to an interchangeable reasoner, thereby discouraging state fabrication and enforcing inference under explicit uncertainty and invariants. We complement this with the Coherence Circle, a low-dimensional geometric interface anchored to an individual baseline attractor. Within this representation, radial displacement indexes drift magnitude, angular sectors capture modes of deviation, and a scalar coherence parameter summarizes stability, cross-modal consistency, and directional alignment over time. The resulting architecture supports closed-loop querying, intervention, and measurement design, while enabling falsifiable evaluation through identity scrambling, modality dropout, trajectory prediction, and intervention-response tests. More broadly, the framework offers an operational basis for studying machine consciousness through interventionally grounded analyses of integration, self-consistency, and active coherence maintenance across biological, ecological, organizational, and artificial systems.