A Unified Language for Cross-AI Model Feature Representation: A Comprehensive Framework for Standardization

This paper presents the Unified Feature Language (UFL), an architecture for intermediate representations that supports cross-model feature exchange and human participation in the evolution of those representations. The central contribution is a single architectural principle: the strict separation of human meaning-making (Cognitive Layer), machine- independent state (Universal Intermediate Representation, UIR), and model-specific trans- lation (Transformation Layer) into three layers with non-overlapping responsibilities.

The previous revision carried the separation to what we argue is its fixed point: the intermediate representation itself must be representation-free. The UIR is defined not as a coordinate space with named dimensions but as a structure identified only up to admissible isomorphism — the invariant of the ε-preserving transformations that connect it to model spaces. Concrete numeric realizations are demoted to charts: versioned, replaceable coor- dinate systems through which all access to state vectors passes. Six design principles force the intermediate unit into the form (id, parent, chart, v): an anonymous vector, well-formed only relative to a declared chart, organized as an append-only tree.

The present revision adds no principle. It adds an empirical hypothesis about what the invariant U contains, prompted by recent evidence that language models maintain inter- nally a small, privileged set of representations with the functional properties of a global workspace (Gurnee et al., 2026). We state the workspace correspondence hypothesis — that U is realized, within useful bounds, by aligning models’ workspace subframes rather than their full representation spaces — together with three falsifiable predictions, a fourth chart-establishing regime that follows from it, and two consequences the architecture must absorb: that a chart inheriting a workspace is a chart whose components arrive already named, which is precisely the condition Chart Neutrality was written to expel; and that the workspace’s documented selectivity implies no single chart can serve both human vocab- ulary evolution and reconstruction-grade fidelity. We regard the hypothesis as the sharpest currently available statement of the architecture’s central empirical dependency, and we are careful to mark where it is conjecture.